Upload folder using huggingface_hub

.env.production

@@ -0,0 +1,29 @@
+# ============================================================
+# OncoAgent — PRODUCTION Environment (AMD Instinct MI300X)
+# Copy to .env on the GPU droplet before deploying.
+# ============================================================
+
+# Set via: export HF_TOKEN=your_token_here
+HF_TOKEN=
+
+# --- Hardware ---
+ROCM_PATH=/opt/rocm
+DEVICE=cuda
+HSA_OVERRIDE_GFX_VERSION=9.4.2
+TENSOR_PARALLEL_SIZE=1
+
+# --- Model Tier IDs ---
+TIER1_MODEL_ID=Qwen/Qwen3.5-9B
+TIER2_MODEL_ID=Qwen/Qwen3.6-27B
+BASE_MODEL_ID=Qwen/Qwen3.5-9B
+
+# --- Inference Backend (Local vLLM) ---
+VLLM_API_BASE=http://localhost:8000/v1
+VLLM_API_KEY=EMPTY
+
+# --- Local LoRA Adapters (MI300X Optimized) ---
+USE_LOCAL_ADAPTERS=false
+LOCAL_ADAPTER_PATH=models/oncoagent_adapters/tier1/checkpoint-1000/
+
+# --- Logging ---
+LOG_LEVEL=INFO

.gitattributes

@@ -33,3 +33,169 @@ saved_model/**/* filter=lfs diff=lfs merge=lfs -text
 *.zip filter=lfs diff=lfs merge=lfs -text
 *.zst filter=lfs diff=lfs merge=lfs -text
 *tfevents* filter=lfs diff=lfs merge=lfs -text
+data/clinical_guides/esmo/ESMO_PMC10416694_Developing_a_core_set_of_patient_reported_outcomes.pdf filter=lfs diff=lfs merge=lfs -text
+data/clinical_guides/esmo/ESMO_PMC10664856_ESMO_ASCO_Recommendations_for_a_Global_Curriculum_.pdf filter=lfs diff=lfs merge=lfs -text
+data/clinical_guides/esmo/ESMO_PMC10774906_ESMO_ASCO_Recommendations_for_a_Global_Curriculum_.pdf filter=lfs diff=lfs merge=lfs -text
+data/clinical_guides/esmo/ESMO_PMC11574484_Effects_of_Baduanjin_exercise_on_cancer_related_fa.pdf filter=lfs diff=lfs merge=lfs -text
+data/clinical_guides/esmo/ESMO_PMC11628549_Research_hotspots_and_trends_in_immunotherapy_for_.pdf filter=lfs diff=lfs merge=lfs -text
+data/clinical_guides/esmo/ESMO_PMC11662070_Characterization_of_shared_neoantigens_landscape_i.pdf filter=lfs diff=lfs merge=lfs -text
+data/clinical_guides/esmo/ESMO_PMC11856526_Multiomics_in_silico_analysis_identifies_TM4SF4_as.pdf filter=lfs diff=lfs merge=lfs -text
+data/clinical_guides/esmo/ESMO_PMC12218492_Plain_language_summary_of_the_THOR_Cohort_1_study_.pdf filter=lfs diff=lfs merge=lfs -text
+data/clinical_guides/esmo/ESMO_PMC12306965_Systematic_critical_appraisal_of_GRADE_recommendat.pdf filter=lfs diff=lfs merge=lfs -text
+data/clinical_guides/esmo/ESMO_PMC12381471_Prevention_and_treatment_of_venous_thromboembolism.pdf filter=lfs diff=lfs merge=lfs -text
+data/clinical_guides/esmo/ESMO_PMC12733557_How_to_Read_a_Next_Generation_Sequencing_Report_fo.pdf filter=lfs diff=lfs merge=lfs -text
+data/clinical_guides/esmo/ESMO_PMC12836719_Adoption_of_electronic_patient_reported_outcomes_i.pdf filter=lfs diff=lfs merge=lfs -text
+data/clinical_guides/esmo/ESMO_PMC12925129_Development_and_formative_evaluation_of_a_follow_u.pdf filter=lfs diff=lfs merge=lfs -text
+data/clinical_guides/esmo/ESMO_PMC12982907_Cost_utility_Analysis_of_R_CHOP_vs_CHOP_in_Patient.pdf filter=lfs diff=lfs merge=lfs -text
+data/clinical_guides/esmo/ESMO_PMC7617288_Randomised_Trial_of_No__Short_term__or_Long_term_A.pdf filter=lfs diff=lfs merge=lfs -text
+data/clinical_guides/esmo/ESMO_PMC8267298_An_evaluation_of_the_reporting_quality_in_clinical.pdf filter=lfs diff=lfs merge=lfs -text
+data/clinical_guides/nccn/Patient[[:space:]]guidelines/Guidelines[[:space:]]for[[:space:]]Detection,[[:space:]]Prevention,[[:space:]]&[[:space:]]Risk[[:space:]]Reduction/breastcancerscreening-patient.pdf filter=lfs diff=lfs merge=lfs -text
+data/clinical_guides/nccn/Patient[[:space:]]guidelines/Guidelines[[:space:]]for[[:space:]]Detection,[[:space:]]Prevention,[[:space:]]&[[:space:]]Risk[[:space:]]Reduction/colorectal-screening-patient.pdf filter=lfs diff=lfs merge=lfs -text
+data/clinical_guides/nccn/Patient[[:space:]]guidelines/Guidelines[[:space:]]for[[:space:]]Detection,[[:space:]]Prevention,[[:space:]]&[[:space:]]Risk[[:space:]]Reduction/genetics-patient.pdf filter=lfs diff=lfs merge=lfs -text
+data/clinical_guides/nccn/Patient[[:space:]]guidelines/Guidelines[[:space:]]for[[:space:]]Detection,[[:space:]]Prevention,[[:space:]]&[[:space:]]Risk[[:space:]]Reduction/lung_screening-patient.pdf filter=lfs diff=lfs merge=lfs -text
+data/clinical_guides/nccn/Patient[[:space:]]guidelines/Guidelines[[:space:]]for[[:space:]]Detection,[[:space:]]Prevention,[[:space:]]&[[:space:]]Risk[[:space:]]Reduction/prostate-screening-patient.pdf filter=lfs diff=lfs merge=lfs -text
+data/clinical_guides/nccn/Patient[[:space:]]guidelines/Guidelines[[:space:]]for[[:space:]]Specific[[:space:]]Populations/aya-patient.pdf filter=lfs diff=lfs merge=lfs -text
+data/clinical_guides/nccn/Patient[[:space:]]guidelines/Guidelines[[:space:]]for[[:space:]]Supportive[[:space:]]Care/GVDH-patient-guideline.pdf filter=lfs diff=lfs merge=lfs -text
+data/clinical_guides/nccn/Patient[[:space:]]guidelines/Guidelines[[:space:]]for[[:space:]]Supportive[[:space:]]Care/bloodclots-patient.pdf filter=lfs diff=lfs merge=lfs -text
+data/clinical_guides/nccn/Patient[[:space:]]guidelines/Guidelines[[:space:]]for[[:space:]]Supportive[[:space:]]Care/distress-patient.pdf filter=lfs diff=lfs merge=lfs -text
+data/clinical_guides/nccn/Patient[[:space:]]guidelines/Guidelines[[:space:]]for[[:space:]]Supportive[[:space:]]Care/fatigue-patient.pdf filter=lfs diff=lfs merge=lfs -text
+data/clinical_guides/nccn/Patient[[:space:]]guidelines/Guidelines[[:space:]]for[[:space:]]Supportive[[:space:]]Care/immunotherapy-checkpoint-patient.pdf filter=lfs diff=lfs merge=lfs -text
+data/clinical_guides/nccn/Patient[[:space:]]guidelines/Guidelines[[:space:]]for[[:space:]]Supportive[[:space:]]Care/immunotherapy-se-car-tcell-patient.pdf filter=lfs diff=lfs merge=lfs -text
+data/clinical_guides/nccn/Patient[[:space:]]guidelines/Guidelines[[:space:]]for[[:space:]]Supportive[[:space:]]Care/low-blood-patient.pdf filter=lfs diff=lfs merge=lfs -text
+data/clinical_guides/nccn/Patient[[:space:]]guidelines/Guidelines[[:space:]]for[[:space:]]Supportive[[:space:]]Care/nausea-patient.pdf filter=lfs diff=lfs merge=lfs -text
+data/clinical_guides/nccn/Patient[[:space:]]guidelines/Guidelines[[:space:]]for[[:space:]]Supportive[[:space:]]Care/palliative-patient.pdf filter=lfs diff=lfs merge=lfs -text
+data/clinical_guides/nccn/Patient[[:space:]]guidelines/Guidelines[[:space:]]for[[:space:]]Supportive[[:space:]]Care/quitting-smoking-patient.pdf filter=lfs diff=lfs merge=lfs -text
+data/clinical_guides/nccn/Patient[[:space:]]guidelines/Guidelines[[:space:]]for[[:space:]]Supportive[[:space:]]Care/survivorship-crl-patient.pdf filter=lfs diff=lfs merge=lfs -text
+data/clinical_guides/nccn/Patient[[:space:]]guidelines/Guidelines[[:space:]]for[[:space:]]Supportive[[:space:]]Care/survivorship-hl-patient.pdf filter=lfs diff=lfs merge=lfs -text
+data/clinical_guides/nccn/Patient[[:space:]]guidelines/Guidelines[[:space:]]for[[:space:]]Treatment[[:space:]]of[[:space:]]Cancer[[:space:]]by[[:space:]]Type/CBCL-patient.pdf filter=lfs diff=lfs merge=lfs -text
+data/clinical_guides/nccn/Patient[[:space:]]guidelines/Guidelines[[:space:]]for[[:space:]]Treatment[[:space:]]of[[:space:]]Cancer[[:space:]]by[[:space:]]Type/Mantle-patient-guideline.pdf filter=lfs diff=lfs merge=lfs -text
+data/clinical_guides/nccn/Patient[[:space:]]guidelines/Guidelines[[:space:]]for[[:space:]]Treatment[[:space:]]of[[:space:]]Cancer[[:space:]]by[[:space:]]Type/PTCL-patient-guideline.pdf filter=lfs diff=lfs merge=lfs -text
+data/clinical_guides/nccn/Patient[[:space:]]guidelines/Guidelines[[:space:]]for[[:space:]]Treatment[[:space:]]of[[:space:]]Cancer[[:space:]]by[[:space:]]Type/SCLC-patient-guideline.pdf filter=lfs diff=lfs merge=lfs -text
+data/clinical_guides/nccn/Patient[[:space:]]guidelines/Guidelines[[:space:]]for[[:space:]]Treatment[[:space:]]of[[:space:]]Cancer[[:space:]]by[[:space:]]Type/adrenal-patient.pdf filter=lfs diff=lfs merge=lfs -text
+data/clinical_guides/nccn/Patient[[:space:]]guidelines/Guidelines[[:space:]]for[[:space:]]Treatment[[:space:]]of[[:space:]]Cancer[[:space:]]by[[:space:]]Type/all-patient.pdf filter=lfs diff=lfs merge=lfs -text
+data/clinical_guides/nccn/Patient[[:space:]]guidelines/Guidelines[[:space:]]for[[:space:]]Treatment[[:space:]]of[[:space:]]Cancer[[:space:]]by[[:space:]]Type/aml-patient.pdf filter=lfs diff=lfs merge=lfs -text
+data/clinical_guides/nccn/Patient[[:space:]]guidelines/Guidelines[[:space:]]for[[:space:]]Treatment[[:space:]]of[[:space:]]Cancer[[:space:]]by[[:space:]]Type/anal-patient.pdf filter=lfs diff=lfs merge=lfs -text
+data/clinical_guides/nccn/Patient[[:space:]]guidelines/Guidelines[[:space:]]for[[:space:]]Treatment[[:space:]]of[[:space:]]Cancer[[:space:]]by[[:space:]]Type/basal-cell-patient-guideline.pdf filter=lfs diff=lfs merge=lfs -text
+data/clinical_guides/nccn/Patient[[:space:]]guidelines/Guidelines[[:space:]]for[[:space:]]Treatment[[:space:]]of[[:space:]]Cancer[[:space:]]by[[:space:]]Type/bladder-patient.pdf filter=lfs diff=lfs merge=lfs -text
+data/clinical_guides/nccn/Patient[[:space:]]guidelines/Guidelines[[:space:]]for[[:space:]]Treatment[[:space:]]of[[:space:]]Cancer[[:space:]]by[[:space:]]Type/bone-patient.pdf filter=lfs diff=lfs merge=lfs -text
+data/clinical_guides/nccn/Patient[[:space:]]guidelines/Guidelines[[:space:]]for[[:space:]]Treatment[[:space:]]of[[:space:]]Cancer[[:space:]]by[[:space:]]Type/brain-gliomas-patient.pdf filter=lfs diff=lfs merge=lfs -text
+data/clinical_guides/nccn/Patient[[:space:]]guidelines/Guidelines[[:space:]]for[[:space:]]Treatment[[:space:]]of[[:space:]]Cancer[[:space:]]by[[:space:]]Type/cervical-patient-guideline.pdf filter=lfs diff=lfs merge=lfs -text
+data/clinical_guides/nccn/Patient[[:space:]]guidelines/Guidelines[[:space:]]for[[:space:]]Treatment[[:space:]]of[[:space:]]Cancer[[:space:]]by[[:space:]]Type/cll-patient.pdf filter=lfs diff=lfs merge=lfs -text
+data/clinical_guides/nccn/Patient[[:space:]]guidelines/Guidelines[[:space:]]for[[:space:]]Treatment[[:space:]]of[[:space:]]Cancer[[:space:]]by[[:space:]]Type/cml-patient.pdf filter=lfs diff=lfs merge=lfs -text
+data/clinical_guides/nccn/Patient[[:space:]]guidelines/Guidelines[[:space:]]for[[:space:]]Treatment[[:space:]]of[[:space:]]Cancer[[:space:]]by[[:space:]]Type/colon-patient.pdf filter=lfs diff=lfs merge=lfs -text
+data/clinical_guides/nccn/Patient[[:space:]]guidelines/Guidelines[[:space:]]for[[:space:]]Treatment[[:space:]]of[[:space:]]Cancer[[:space:]]by[[:space:]]Type/ctcl-patient.pdf filter=lfs diff=lfs merge=lfs -text
+data/clinical_guides/nccn/Patient[[:space:]]guidelines/Guidelines[[:space:]]for[[:space:]]Treatment[[:space:]]of[[:space:]]Cancer[[:space:]]by[[:space:]]Type/esophageal-patient.pdf filter=lfs diff=lfs merge=lfs -text
+data/clinical_guides/nccn/Patient[[:space:]]guidelines/Guidelines[[:space:]]for[[:space:]]Treatment[[:space:]]of[[:space:]]Cancer[[:space:]]by[[:space:]]Type/gallandbile-hp-patient.pdf filter=lfs diff=lfs merge=lfs -text
+data/clinical_guides/nccn/Patient[[:space:]]guidelines/Guidelines[[:space:]]for[[:space:]]Treatment[[:space:]]of[[:space:]]Cancer[[:space:]]by[[:space:]]Type/gist-patient.pdf filter=lfs diff=lfs merge=lfs -text
+data/clinical_guides/nccn/Patient[[:space:]]guidelines/Guidelines[[:space:]]for[[:space:]]Treatment[[:space:]]of[[:space:]]Cancer[[:space:]]by[[:space:]]Type/hn-nasopharynx-patient.pdf filter=lfs diff=lfs merge=lfs -text
+data/clinical_guides/nccn/Patient[[:space:]]guidelines/Guidelines[[:space:]]for[[:space:]]Treatment[[:space:]]of[[:space:]]Cancer[[:space:]]by[[:space:]]Type/hn-oral-patient.pdf filter=lfs diff=lfs merge=lfs -text
+data/clinical_guides/nccn/Patient[[:space:]]guidelines/Guidelines[[:space:]]for[[:space:]]Treatment[[:space:]]of[[:space:]]Cancer[[:space:]]by[[:space:]]Type/hn-oropharyngeal-patient.pdf filter=lfs diff=lfs merge=lfs -text
+data/clinical_guides/nccn/Patient[[:space:]]guidelines/Guidelines[[:space:]]for[[:space:]]Treatment[[:space:]]of[[:space:]]Cancer[[:space:]]by[[:space:]]Type/hodgkin-patient.pdf filter=lfs diff=lfs merge=lfs -text
+data/clinical_guides/nccn/Patient[[:space:]]guidelines/Guidelines[[:space:]]for[[:space:]]Treatment[[:space:]]of[[:space:]]Cancer[[:space:]]by[[:space:]]Type/hodgkinlymphomainchildren-patient.pdf filter=lfs diff=lfs merge=lfs -text
+data/clinical_guides/nccn/Patient[[:space:]]guidelines/Guidelines[[:space:]]for[[:space:]]Treatment[[:space:]]of[[:space:]]Cancer[[:space:]]by[[:space:]]Type/inflammatory-breast-patient.pdf filter=lfs diff=lfs merge=lfs -text
+data/clinical_guides/nccn/Patient[[:space:]]guidelines/Guidelines[[:space:]]for[[:space:]]Treatment[[:space:]]of[[:space:]]Cancer[[:space:]]by[[:space:]]Type/kidney-patient.pdf filter=lfs diff=lfs merge=lfs -text
+data/clinical_guides/nccn/Patient[[:space:]]guidelines/Guidelines[[:space:]]for[[:space:]]Treatment[[:space:]]of[[:space:]]Cancer[[:space:]]by[[:space:]]Type/liver-hp-patient.pdf filter=lfs diff=lfs merge=lfs -text
+data/clinical_guides/nccn/Patient[[:space:]]guidelines/Guidelines[[:space:]]for[[:space:]]Treatment[[:space:]]of[[:space:]]Cancer[[:space:]]by[[:space:]]Type/lung-early-stage-patient.pdf filter=lfs diff=lfs merge=lfs -text
+data/clinical_guides/nccn/Patient[[:space:]]guidelines/Guidelines[[:space:]]for[[:space:]]Treatment[[:space:]]of[[:space:]]Cancer[[:space:]]by[[:space:]]Type/lung-metastatic-patient.pdf filter=lfs diff=lfs merge=lfs -text
+data/clinical_guides/nccn/Patient[[:space:]]guidelines/Guidelines[[:space:]]for[[:space:]]Treatment[[:space:]]of[[:space:]]Cancer[[:space:]]by[[:space:]]Type/mds-patient.pdf filter=lfs diff=lfs merge=lfs -text
+data/clinical_guides/nccn/Patient[[:space:]]guidelines/Guidelines[[:space:]]for[[:space:]]Treatment[[:space:]]of[[:space:]]Cancer[[:space:]]by[[:space:]]Type/melanoma-patient.pdf filter=lfs diff=lfs merge=lfs -text
+data/clinical_guides/nccn/Patient[[:space:]]guidelines/Guidelines[[:space:]]for[[:space:]]Treatment[[:space:]]of[[:space:]]Cancer[[:space:]]by[[:space:]]Type/mpm-patient.pdf filter=lfs diff=lfs merge=lfs -text
+data/clinical_guides/nccn/Patient[[:space:]]guidelines/Guidelines[[:space:]]for[[:space:]]Treatment[[:space:]]of[[:space:]]Cancer[[:space:]]by[[:space:]]Type/mpn-patient.pdf filter=lfs diff=lfs merge=lfs -text
+data/clinical_guides/nccn/Patient[[:space:]]guidelines/Guidelines[[:space:]]for[[:space:]]Treatment[[:space:]]of[[:space:]]Cancer[[:space:]]by[[:space:]]Type/myeloma-patient.pdf filter=lfs diff=lfs merge=lfs -text
+data/clinical_guides/nccn/Patient[[:space:]]guidelines/Guidelines[[:space:]]for[[:space:]]Treatment[[:space:]]of[[:space:]]Cancer[[:space:]]by[[:space:]]Type/mzl-patient.pdf filter=lfs diff=lfs merge=lfs -text
+data/clinical_guides/nccn/Patient[[:space:]]guidelines/Guidelines[[:space:]]for[[:space:]]Treatment[[:space:]]of[[:space:]]Cancer[[:space:]]by[[:space:]]Type/neuroendocrine-patient.pdf filter=lfs diff=lfs merge=lfs -text
+data/clinical_guides/nccn/Patient[[:space:]]guidelines/Guidelines[[:space:]]for[[:space:]]Treatment[[:space:]]of[[:space:]]Cancer[[:space:]]by[[:space:]]Type/nhl-diffuse-patient.pdf filter=lfs diff=lfs merge=lfs -text
+data/clinical_guides/nccn/Patient[[:space:]]guidelines/Guidelines[[:space:]]for[[:space:]]Treatment[[:space:]]of[[:space:]]Cancer[[:space:]]by[[:space:]]Type/nhl-follicular-patient.pdf filter=lfs diff=lfs merge=lfs -text
+data/clinical_guides/nccn/Patient[[:space:]]guidelines/Guidelines[[:space:]]for[[:space:]]Treatment[[:space:]]of[[:space:]]Cancer[[:space:]]by[[:space:]]Type/ovarian-patient.pdf filter=lfs diff=lfs merge=lfs -text
+data/clinical_guides/nccn/Patient[[:space:]]guidelines/Guidelines[[:space:]]for[[:space:]]Treatment[[:space:]]of[[:space:]]Cancer[[:space:]]by[[:space:]]Type/pancreatic-patient.pdf filter=lfs diff=lfs merge=lfs -text
+data/clinical_guides/nccn/Patient[[:space:]]guidelines/Guidelines[[:space:]]for[[:space:]]Treatment[[:space:]]of[[:space:]]Cancer[[:space:]]by[[:space:]]Type/pcnsl-patient.pdf filter=lfs diff=lfs merge=lfs -text
+data/clinical_guides/nccn/Patient[[:space:]]guidelines/Guidelines[[:space:]]for[[:space:]]Treatment[[:space:]]of[[:space:]]Cancer[[:space:]]by[[:space:]]Type/ped_all_patient.pdf filter=lfs diff=lfs merge=lfs -text
+data/clinical_guides/nccn/Patient[[:space:]]guidelines/Guidelines[[:space:]]for[[:space:]]Treatment[[:space:]]of[[:space:]]Cancer[[:space:]]by[[:space:]]Type/prostate-advanced-patient.pdf filter=lfs diff=lfs merge=lfs -text
+data/clinical_guides/nccn/Patient[[:space:]]guidelines/Guidelines[[:space:]]for[[:space:]]Treatment[[:space:]]of[[:space:]]Cancer[[:space:]]by[[:space:]]Type/prostate-early-patient.pdf filter=lfs diff=lfs merge=lfs -text
+data/clinical_guides/nccn/Patient[[:space:]]guidelines/Guidelines[[:space:]]for[[:space:]]Treatment[[:space:]]of[[:space:]]Cancer[[:space:]]by[[:space:]]Type/rectal-patient.pdf filter=lfs diff=lfs merge=lfs -text
+data/clinical_guides/nccn/Patient[[:space:]]guidelines/Guidelines[[:space:]]for[[:space:]]Treatment[[:space:]]of[[:space:]]Cancer[[:space:]]by[[:space:]]Type/sarcoma-patient.pdf filter=lfs diff=lfs merge=lfs -text
+data/clinical_guides/nccn/Patient[[:space:]]guidelines/Guidelines[[:space:]]for[[:space:]]Treatment[[:space:]]of[[:space:]]Cancer[[:space:]]by[[:space:]]Type/small-bowel-patient.pdf filter=lfs diff=lfs merge=lfs -text
+data/clinical_guides/nccn/Patient[[:space:]]guidelines/Guidelines[[:space:]]for[[:space:]]Treatment[[:space:]]of[[:space:]]Cancer[[:space:]]by[[:space:]]Type/squamous_cell-patient.pdf filter=lfs diff=lfs merge=lfs -text
+data/clinical_guides/nccn/Patient[[:space:]]guidelines/Guidelines[[:space:]]for[[:space:]]Treatment[[:space:]]of[[:space:]]Cancer[[:space:]]by[[:space:]]Type/stage_0_breast-patient.pdf filter=lfs diff=lfs merge=lfs -text
+data/clinical_guides/nccn/Patient[[:space:]]guidelines/Guidelines[[:space:]]for[[:space:]]Treatment[[:space:]]of[[:space:]]Cancer[[:space:]]by[[:space:]]Type/stage_iv_breast-patient.pdf filter=lfs diff=lfs merge=lfs -text
+data/clinical_guides/nccn/Patient[[:space:]]guidelines/Guidelines[[:space:]]for[[:space:]]Treatment[[:space:]]of[[:space:]]Cancer[[:space:]]by[[:space:]]Type/stomach-patient.pdf filter=lfs diff=lfs merge=lfs -text
+data/clinical_guides/nccn/Patient[[:space:]]guidelines/Guidelines[[:space:]]for[[:space:]]Treatment[[:space:]]of[[:space:]]Cancer[[:space:]]by[[:space:]]Type/systemic-mastocytosis-patient-guideline.pdf filter=lfs diff=lfs merge=lfs -text
+data/clinical_guides/nccn/Patient[[:space:]]guidelines/Guidelines[[:space:]]for[[:space:]]Treatment[[:space:]]of[[:space:]]Cancer[[:space:]]by[[:space:]]Type/thyroid-patient.pdf filter=lfs diff=lfs merge=lfs -text
+data/clinical_guides/nccn/Patient[[:space:]]guidelines/Guidelines[[:space:]]for[[:space:]]Treatment[[:space:]]of[[:space:]]Cancer[[:space:]]by[[:space:]]Type/uterine-patient.pdf filter=lfs diff=lfs merge=lfs -text
+data/clinical_guides/nccn/Patient[[:space:]]guidelines/Guidelines[[:space:]]for[[:space:]]Treatment[[:space:]]of[[:space:]]Cancer[[:space:]]by[[:space:]]Type/waldenstrom-patient.pdf filter=lfs diff=lfs merge=lfs -text
+data/clinical_guides/nccn/Professional[[:space:]]guidelines/aml.pdf filter=lfs diff=lfs merge=lfs -text
+data/clinical_guides/nccn/Professional[[:space:]]guidelines/ampullary.pdf filter=lfs diff=lfs merge=lfs -text
+data/clinical_guides/nccn/Professional[[:space:]]guidelines/amyloidosis.pdf filter=lfs diff=lfs merge=lfs -text
+data/clinical_guides/nccn/Professional[[:space:]]guidelines/anal.pdf filter=lfs diff=lfs merge=lfs -text
+data/clinical_guides/nccn/Professional[[:space:]]guidelines/appendiceal.pdf filter=lfs diff=lfs merge=lfs -text
+data/clinical_guides/nccn/Professional[[:space:]]guidelines/b-cell.pdf filter=lfs diff=lfs merge=lfs -text
+data/clinical_guides/nccn/Professional[[:space:]]guidelines/bladder.pdf filter=lfs diff=lfs merge=lfs -text
+data/clinical_guides/nccn/Professional[[:space:]]guidelines/bone.pdf filter=lfs diff=lfs merge=lfs -text
+data/clinical_guides/nccn/Professional[[:space:]]guidelines/breast.pdf filter=lfs diff=lfs merge=lfs -text
+data/clinical_guides/nccn/Professional[[:space:]]guidelines/btc.pdf filter=lfs diff=lfs merge=lfs -text
+data/clinical_guides/nccn/Professional[[:space:]]guidelines/castleman.pdf filter=lfs diff=lfs merge=lfs -text
+data/clinical_guides/nccn/Professional[[:space:]]guidelines/cervical.pdf filter=lfs diff=lfs merge=lfs -text
+data/clinical_guides/nccn/Professional[[:space:]]guidelines/cll.pdf filter=lfs diff=lfs merge=lfs -text
+data/clinical_guides/nccn/Professional[[:space:]]guidelines/cml.pdf filter=lfs diff=lfs merge=lfs -text
+data/clinical_guides/nccn/Professional[[:space:]]guidelines/cns.pdf filter=lfs diff=lfs merge=lfs -text
+data/clinical_guides/nccn/Professional[[:space:]]guidelines/colon.pdf filter=lfs diff=lfs merge=lfs -text
+data/clinical_guides/nccn/Professional[[:space:]]guidelines/cutaneous_lymphomas.pdf filter=lfs diff=lfs merge=lfs -text
+data/clinical_guides/nccn/Professional[[:space:]]guidelines/cutaneous_melanoma.pdf filter=lfs diff=lfs merge=lfs -text
+data/clinical_guides/nccn/Professional[[:space:]]guidelines/dfsp.pdf filter=lfs diff=lfs merge=lfs -text
+data/clinical_guides/nccn/Professional[[:space:]]guidelines/esophageal.pdf filter=lfs diff=lfs merge=lfs -text
+data/clinical_guides/nccn/Professional[[:space:]]guidelines/gastric.pdf filter=lfs diff=lfs merge=lfs -text
+data/clinical_guides/nccn/Professional[[:space:]]guidelines/gist.pdf filter=lfs diff=lfs merge=lfs -text
+data/clinical_guides/nccn/Professional[[:space:]]guidelines/gtn.pdf filter=lfs diff=lfs merge=lfs -text
+data/clinical_guides/nccn/Professional[[:space:]]guidelines/hairy_cell.pdf filter=lfs diff=lfs merge=lfs -text
+data/clinical_guides/nccn/Professional[[:space:]]guidelines/hcc.pdf filter=lfs diff=lfs merge=lfs -text
+data/clinical_guides/nccn/Professional[[:space:]]guidelines/head-and-neck.pdf filter=lfs diff=lfs merge=lfs -text
+data/clinical_guides/nccn/Professional[[:space:]]guidelines/histiocytic_neoplasms.pdf filter=lfs diff=lfs merge=lfs -text
+data/clinical_guides/nccn/Professional[[:space:]]guidelines/hodgkins.pdf filter=lfs diff=lfs merge=lfs -text
+data/clinical_guides/nccn/Professional[[:space:]]guidelines/immunotherapy_infographic.pdf filter=lfs diff=lfs merge=lfs -text
+data/clinical_guides/nccn/Professional[[:space:]]guidelines/kaposi.pdf filter=lfs diff=lfs merge=lfs -text
+data/clinical_guides/nccn/Professional[[:space:]]guidelines/kidney.pdf filter=lfs diff=lfs merge=lfs -text
+data/clinical_guides/nccn/Professional[[:space:]]guidelines/mastocytosis.pdf filter=lfs diff=lfs merge=lfs -text
+data/clinical_guides/nccn/Professional[[:space:]]guidelines/mcc.pdf filter=lfs diff=lfs merge=lfs -text
+data/clinical_guides/nccn/Professional[[:space:]]guidelines/mds.pdf filter=lfs diff=lfs merge=lfs -text
+data/clinical_guides/nccn/Professional[[:space:]]guidelines/meso_peritoneal.pdf filter=lfs diff=lfs merge=lfs -text
+data/clinical_guides/nccn/Professional[[:space:]]guidelines/meso_pleural.pdf filter=lfs diff=lfs merge=lfs -text
+data/clinical_guides/nccn/Professional[[:space:]]guidelines/mlne.pdf filter=lfs diff=lfs merge=lfs -text
+data/clinical_guides/nccn/Professional[[:space:]]guidelines/mpn.pdf filter=lfs diff=lfs merge=lfs -text
+data/clinical_guides/nccn/Professional[[:space:]]guidelines/myeloma.pdf filter=lfs diff=lfs merge=lfs -text
+data/clinical_guides/nccn/Professional[[:space:]]guidelines/neuroblastoma.pdf filter=lfs diff=lfs merge=lfs -text
+data/clinical_guides/nccn/Professional[[:space:]]guidelines/neuroendocrine.pdf filter=lfs diff=lfs merge=lfs -text
+data/clinical_guides/nccn/Professional[[:space:]]guidelines/nmsc.pdf filter=lfs diff=lfs merge=lfs -text
+data/clinical_guides/nccn/Professional[[:space:]]guidelines/nscl.pdf filter=lfs diff=lfs merge=lfs -text
+data/clinical_guides/nccn/Professional[[:space:]]guidelines/occult.pdf filter=lfs diff=lfs merge=lfs -text
+data/clinical_guides/nccn/Professional[[:space:]]guidelines/ovarian.pdf filter=lfs diff=lfs merge=lfs -text
+data/clinical_guides/nccn/Professional[[:space:]]guidelines/pancreatic.pdf filter=lfs diff=lfs merge=lfs -text
+data/clinical_guides/nccn/Professional[[:space:]]guidelines/ped_all.pdf filter=lfs diff=lfs merge=lfs -text
+data/clinical_guides/nccn/Professional[[:space:]]guidelines/ped_b-cell.pdf filter=lfs diff=lfs merge=lfs -text
+data/clinical_guides/nccn/Professional[[:space:]]guidelines/ped_cns.pdf filter=lfs diff=lfs merge=lfs -text
+data/clinical_guides/nccn/Professional[[:space:]]guidelines/ped_hodgkin.pdf filter=lfs diff=lfs merge=lfs -text
+data/clinical_guides/nccn/Professional[[:space:]]guidelines/ped_sts.pdf filter=lfs diff=lfs merge=lfs -text
+data/clinical_guides/nccn/Professional[[:space:]]guidelines/penile.pdf filter=lfs diff=lfs merge=lfs -text
+data/clinical_guides/nccn/Professional[[:space:]]guidelines/prostate.pdf filter=lfs diff=lfs merge=lfs -text
+data/clinical_guides/nccn/Professional[[:space:]]guidelines/rectal.pdf filter=lfs diff=lfs merge=lfs -text
+data/clinical_guides/nccn/Professional[[:space:]]guidelines/sarcoma.pdf filter=lfs diff=lfs merge=lfs -text
+data/clinical_guides/nccn/Professional[[:space:]]guidelines/sclc.pdf filter=lfs diff=lfs merge=lfs -text
+data/clinical_guides/nccn/Professional[[:space:]]guidelines/small_bowel.pdf filter=lfs diff=lfs merge=lfs -text
+data/clinical_guides/nccn/Professional[[:space:]]guidelines/squamous.pdf filter=lfs diff=lfs merge=lfs -text
+data/clinical_guides/nccn/Professional[[:space:]]guidelines/t-cell.pdf filter=lfs diff=lfs merge=lfs -text
+data/clinical_guides/nccn/Professional[[:space:]]guidelines/testicular.pdf filter=lfs diff=lfs merge=lfs -text
+data/clinical_guides/nccn/Professional[[:space:]]guidelines/thymic.pdf filter=lfs diff=lfs merge=lfs -text
+data/clinical_guides/nccn/Professional[[:space:]]guidelines/thyroid.pdf filter=lfs diff=lfs merge=lfs -text
+data/clinical_guides/nccn/Professional[[:space:]]guidelines/uterine.pdf filter=lfs diff=lfs merge=lfs -text
+data/clinical_guides/nccn/Professional[[:space:]]guidelines/uveal.pdf filter=lfs diff=lfs merge=lfs -text
+data/clinical_guides/nccn/Professional[[:space:]]guidelines/vaginal.pdf filter=lfs diff=lfs merge=lfs -text
+data/clinical_guides/nccn/Professional[[:space:]]guidelines/vulvar.pdf filter=lfs diff=lfs merge=lfs -text
+data/clinical_guides/nccn/Professional[[:space:]]guidelines/waldenstroms.pdf filter=lfs diff=lfs merge=lfs -text
+data/clinical_guides/nccn/Professional[[:space:]]guidelines/wilms_tumor.pdf filter=lfs diff=lfs merge=lfs -text
+data/clinical_guides/nccn/immunotherapy_infographic.pdf filter=lfs diff=lfs merge=lfs -text
+data/clinical_guides/nccn/nccn_distress_thermometer.pdf filter=lfs diff=lfs merge=lfs -text
+data/clinical_guides/nccn/questions-to-ask-about-cancer-care.pdf filter=lfs diff=lfs merge=lfs -text
+docs/OncoAgent_Official_Paper.pdf filter=lfs diff=lfs merge=lfs -text
+docs/assets/brand/colors/color_palette.png filter=lfs diff=lfs merge=lfs -text
+docs/assets/brand/logo/oncoagent_logo_dark_mode.png filter=lfs diff=lfs merge=lfs -text
+docs/assets/brand/logo/oncoagent_logo_full_color.png filter=lfs diff=lfs merge=lfs -text
+docs/assets/brand/social/twitter_header.png filter=lfs diff=lfs merge=lfs -text

Dockerfile

@@ -0,0 +1,79 @@
+# ============================================================
+# OncoAgent — Production Dockerfile
+# Hardware: AMD Instinct MI300X / ROCm 7.2
+# Serves: vLLM (Qwen3.5-9B + Qwen3.6-27B) + Gradio UI
+# ============================================================
+
+# Base image: vLLM optimized for ROCm
+FROM rocm/vllm:latest
+
+# System environment
+ENV DEBIAN_FRONTEND=noninteractive
+ENV PYTHONUNBUFFERED=1
+ENV GRADIO_SERVER_NAME="0.0.0.0"
+ENV GRADIO_SERVER_PORT=7860
+
+# ROCm / PyTorch environment
+ENV HSA_OVERRIDE_GFX_VERSION=9.4.2
+ENV PYTORCH_ROCM_ARCH="gfx942"
+
+# OncoAgent model configuration
+ENV TIER1_MODEL_ID="Qwen/Qwen3.5-9B"
+ENV TIER2_MODEL_ID="Qwen/Qwen3.6-27B"
+ENV BASE_MODEL_ID="Qwen/Qwen3.5-9B"
+ENV VLLM_API_BASE="http://localhost:8000/v1"
+ENV VLLM_API_KEY="EMPTY"
+ENV USE_LOCAL_ADAPTERS="false"
+ENV DEVICE="cuda"
+ENV TENSOR_PARALLEL_SIZE=1
+
+WORKDIR /app
+
+# System dependencies
+RUN apt-get update && apt-get install -y --no-install-recommends \
+    git \
+    build-essential \
+    supervisor \
+    && rm -rf /var/lib/apt/lists/*
+
+# Python dependencies
+COPY requirements.txt /app/
+RUN pip install --no-cache-dir -r requirements.txt
+
+# Application code
+COPY . /app/
+
+# Make deploy scripts executable
+RUN chmod +x deploy/start_vllm.sh
+
+# Supervisor config to run vLLM + Gradio simultaneously
+RUN cat > /etc/supervisor/conf.d/oncoagent.conf <<'EOF'
+[supervisord]
+nodaemon=true
+logfile=/var/log/supervisord.log
+
+[program:vllm]
+command=bash /app/deploy/start_vllm.sh tier1
+directory=/app
+autostart=true
+autorestart=true
+stdout_logfile=/var/log/vllm.log
+stderr_logfile=/var/log/vllm_err.log
+priority=10
+
+[program:gradio]
+command=python /app/ui/app.py
+directory=/app
+autostart=true
+autorestart=true
+stdout_logfile=/var/log/gradio.log
+stderr_logfile=/var/log/gradio_err.log
+priority=20
+startsecs=30
+EOF
+
+# Expose ports: Gradio (7860) + vLLM API (8000)
+EXPOSE 7860 8000
+
+# Start both services via supervisor
+CMD ["supervisord", "-c", "/etc/supervisor/conf.d/oncoagent.conf"]

README.es.md

@@ -0,0 +1,129 @@
+# 🧬 OncoAgent — Sistema Multi-Agente de Triaje Oncológico
+
+![ROCm](https://img.shields.io/badge/AMD-ROCm_7.2-ed1c24?logo=amd&logoColor=white)
+![Python](https://img.shields.io/badge/Python-3.10+-3776AB?logo=python&logoColor=white)
+![vLLM](https://img.shields.io/badge/vLLM-PagedAttention-000000?logo=vllm&logoColor=white)
+![LangGraph](https://img.shields.io/badge/Orchestration-LangGraph-FF4F00?logo=langchain&logoColor=white)
+![Gradio](https://img.shields.io/badge/UI-Gradio_6-FF7C00?logo=gradio&logoColor=white)
+
+> **AMD Developer Hackathon 2026** · Potenciado por AMD Instinct™ MI300X · ROCm 7.2
+
+## 🌍 100% Código Abierto: Democratizando la Oncología
+OncoAgent es orgullosamente 100% de código abierto. Creemos que la inteligencia clínica capaz de salvar vidas no debería estar bloqueada tras APIs propietarias. Nuestra solución está diseñada para:
+- **Garantizar la Privacidad del Paciente:** Ejecutarse localmente en hardware AMD MI300X o nubes privadas, asegurando que ningún dato del paciente abandone el hospital.
+- **Fomentar la Contribución Global:** Permitir a las comunidades médicas de todo el mundo auditar, modificar y contribuir fácilmente a la base de conocimiento RAG.
+
+OncoAgent es un sistema de triaje clínico multi-agente de última generación diseñado para combatir la **ceguera por datos no estructurados** en la oncología de atención primaria. Aprovecha una arquitectura adaptativa por niveles con modelos **Qwen 3.5-9B** (Triaje Rápido) y **Qwen 3.6-27B** (Razonamiento Profundo). Orquestado a través de una sofisticada máquina de estados de LangGraph, proporciona razonamiento oncológico basado en evidencia estrictamente fundamentado en las guías clínicas de NCCN/ESMO, con puertas de seguridad de validación humana (HITL) integradas y un bucle de crítica basado en Reflexion.
+
+---
+
+## 🏗️ Arquitectura
+
+```
+┌────────┐   ┌─────────┐   ┌─────────┐   ┌────────────┐      ┌────────────┐   ┌─────────┐
+│ Router │──▶│Ingestión│──▶│   RAG   │──▶│Especialista│◀────│  Crítico   │   │Formateo │
+│(Triaje)│   │ (PHI)   │   │Correctiv│   │ (Qwen 9B/  │     │(Validación │   │(Salida) │
+└────────┘   └─────────┘   └─────────┘   │    27B)    │────▶│ Reflexion) │   └─────────┘
+    │           │             │          └────────────┘      └────────────┘        ▲
+    │           │             │                 │                   │              │
+    ▼           ▼             ▼                 ▼                   ▼              │
+  ┌───────────────────────────────────────────────────────────────────┐      ┌────────────┐
+  │                        Nodo de Respaldo (Fallback)                │      │Puerta HITL │
+  └───────────────────────────────────────────────────────────────────┘      │(Agudeza)   │
+                                                                             └────────────┘
+```
+
+**Componentes Principales:**
+
+| Módulo | Descripción |
+|--------|-------------|
+| `data_prep/` | Constructor del dataset: PMC-Patients/OncoCoT → Strict JSONL (Plantilla chat de Llama 3) |
+| `rag_engine/` | El "Cerebro": Extracción PyMuPDF, Semantic Chunking Adaptativo de PDFs NCCN/ESMO, & Vectorización ChromaDB + PubMedBERT. |
+| `agents/` | El "Razonamiento": Orquestación multi-agente LangGraph (Router → Corrective RAG → Specialist ↔ Critic → HITL Gate). |
+| `ui/` | La "Cara": Interfaz Gradio 6 con Glassmorphism para input clínico, citas en tiempo real y salida estructurada. |
+
+---
+
+## 🧠 Estrategia de Modelo Dual-Tier (Qwen)
+
+Para maximizar las capacidades de cómputo del **AMD MI300X**, OncoAgent implementa una estrategia de enrutamiento dinámica de **Doble Nivel (Dual-Tier)** utilizando la familia de modelos Qwen. **Ambos niveles (tiers) han sido ajustados (fine-tuned) en más de 200,000 casos oncológicos clínicos reales, cubriendo todos los tipos principales de cáncer** (derivados de los datasets PMC-Patients y OncoCoT) para garantizar un razonamiento médico hiper-especializado:
+
+- **Tier 1: Qwen 3.5-9B (Speed Triage):** Un modelo extremadamente rápido y ligero usado por el `Router` para evaluar la complejidad inicial, realizar triaje simple y procesar consultas de bajo riesgo.
+- **Tier 2: Qwen 3.6-27B (Deep Reasoning):** El modelo pesado. Se activa para casos clínicos de alta complejidad (ej. metástasis, mutaciones múltiples). Realiza un razonamiento profundo y verificaciones de entrelazamiento (entailment checks), evitando el sesgo de confirmación mediante rigurosos bucles de Reflexion.
+
+---
+
+## ⚡ Objetivo de Hardware
+
+- **GPU:** AMD Instinct™ MI300X (192GB HBM3)
+- **Pila de Software:** ROCm 7.2.x, PyTorch (HIP), vLLM con PagedAttention
+- **Modelos:** `Qwen/Qwen3.5-9B` (Triaje Rápido) y `Qwen/Qwen3.6-27B-Instruct` (Razonamiento Profundo)
+- **Precisión:** QLoRA 4-bit NormalFloat4 vía `bitsandbytes` (Compatible con ROCm)
+
+---
+
+## 🚀 Inicio Rápido
+
+```bash
+# 1. Clonar y configurar
+git clone <repo-url>
+cd OncoAgent
+
+# 2. Instalar dependencias
+python -m venv .venv
+source .venv/bin/activate
+pip install -r requirements.txt
+
+# 3. Iniciar Servidor de Inferencia (vLLM en Docker)
+# Esto levanta los modelos Qwen optimizados para AMD MI300X vía ROCm PagedAttention
+docker run --device /dev/kfd --device /dev/dri -p 8000:8000 rocm/vllm:latest \
+    --model Qwen/Qwen3.6-27B-Instruct --tensor-parallel-size 1
+
+# 4. Configurar entorno y ejecutar interfaz
+cp .env.example .env
+# Configurar VLLM_API_BASE=http://localhost:8000/v1 en .env
+python -m ui.app
+```
+
+---
+
+## 📁 Estructura del Proyecto
+
+```
+├── docs/                   # Documentación e investigación
+│   ├── research/           # Documentos de análisis de investigación profunda
+│   ├── ADR/                # Registros de Decisiones Arquitectónicas (ADRs)
+│   ├── oncoagent_master_directive.md
+│   └── antigravity_rules.md
+├── data_prep/              # Preparación de conjuntos de datos (Fase 0)
+├── rag_engine/             # Ingestión y recuperación de RAG (Fase 0-3)
+├── agents/                 # Orquestación LangGraph (Fase 3)
+├── ui/                     # Frontend en Gradio (Fase 4)
+├── tests/                  # Pruebas unitarias e integración
+├── scripts/                # Scripts de utilidad
+├── logs/                   # Bitácora (Paper log) y de redes sociales
+├── requirements.txt        # Dependencias fijadas
+└── Dockerfile              # Despliegue en HF Spaces
+```
+
+---
+
+## 🩺 Garantías de Seguridad
+
+- **Bucle Crítico basado en Reflexion:** Un nodo de seguridad dedicado audita la salida del Especialista contra el contexto RAG (verificación de implicación). Obliga al Especialista a regenerar su salida si detecta afirmaciones sin fundamento o dosis inventadas.
+- **Puerta de Aprobación Humana (HITL):** Un punto de control basado en la agudeza clínica que detiene el flujo para la aprobación de un médico humano en casos de alto riesgo (ej. Estadio IV + mutaciones complejas).
+- **RAG Correctivo:** El sistema evalúa la relevancia del contexto recuperado. Si no se encuentra evidencia suficiente, se activa un respaldo seguro en lugar de intentar adivinar.
+- **Cero-PHI (Cero Información Médica Privada):** Redacción de PII basada en expresiones regulares antes de cualquier procesamiento.
+- **Reproducibilidad:** Semillas fijas (`torch.manual_seed(42)`) en todos los scripts de ML.
+
+---
+
+## 📄 Licencia
+
+Este proyecto fue construido para el AMD Developer Hackathon 2026.
+
+---
+
+## 👥 Equipo
+
+Construido con ❤️ y AMD Instinct MI300X.

README.md

@@ -6,9 +6,9 @@ colorTo: blue
 sdk: gradio
 sdk_version: 5.31.0
 app_file: app.py
-pinned: true
+pinned: false
 license: apache-2.0
-short_description: Multi-agent oncology AI saving lives ¡fast cancer detection!
+short_description: Multi-Agent Oncology Triage powered by AMD MI300X
 ---
 
 # 🧬 OncoAgent — Multi-Agent Oncology Triage System
@@ -139,4 +139,4 @@ This project was built for the AMD Developer Hackathon 2026.
 
 ## 👥 Team
 
-Built with ❤️ and AMD Instinct MI300X.
+Built with ❤️ and AMD Instinct MI300X.

agents/__init__.py

@@ -0,0 +1,32 @@
+"""
+OncoAgent Multi-Agent System — SOTA Architecture.
+
+This package implements a production-grade clinical oncology triage system
+using LangGraph for orchestration, incorporating:
+
+  - Router: complexity classification + model tier selection
+  - Corrective RAG: graded retrieval with query rewriting
+  - Specialist: tier-adaptive clinical reasoning (9B/27B)
+  - Critic: reflexion-pattern validation loop
+  - HITL Gate: clinician approval for high-acuity cases
+  - Formatter: structured output with confidence metrics
+
+Architecture inspired by Claude Code, Hermes Agent, Corrective RAG,
+and Reflexion patterns.
+"""
+
+from .graph import build_oncoagent_graph
+from .state import AgentState
+from .memory import get_memory_store, PatientMemoryStore
+from .tools import get_vllm_client, call_tier_model, get_tier_spec, TIER_SPECS
+
+__all__ = [
+    "build_oncoagent_graph",
+    "AgentState",
+    "get_memory_store",
+    "PatientMemoryStore",
+    "get_vllm_client",
+    "call_tier_model",
+    "get_tier_spec",
+    "TIER_SPECS",
+]

agents/corrective_rag.py

@@ -0,0 +1,353 @@
+"""
+Corrective RAG Node — Graded retrieval with query rewriting.
+
+Design pattern: Corrective RAG (CRAG) from Yan et al. 2024
+  1. Retrieve top-K documents from ChromaDB
+  2. Grade each document for relevance (binary: RELEVANT / IRRELEVANT)
+  3. If insufficient relevant docs → rewrite query and re-retrieve
+  4. If still insufficient after max retries → route to fallback
+
+Also implements parallelised evidence gathering:
+  - ChromaDB (clinical guidelines)
+  - CIViC API (genomic evidence)
+  - ClinicalTrials.gov (active trials)
+"""
+
+import logging
+import re
+from typing import Dict, Any, List, Optional
+from concurrent.futures import ThreadPoolExecutor, as_completed
+
+from .state import AgentState
+from .tools import call_tier_model
+
+logger = logging.getLogger(__name__)
+
+
+# ---------------------------------------------------------------------------
+# Lazy-loaded retriever singleton
+# ---------------------------------------------------------------------------
+
+_retriever_instance = None
+
+
+def _get_retriever():
+    """Return a cached OncoRAGRetriever instance (lazy init)."""
+    global _retriever_instance
+    if _retriever_instance is None:
+        try:
+            from rag_engine.retriever import OncoRAGRetriever
+            _retriever_instance = OncoRAGRetriever()
+            logger.info("OncoRAGRetriever initialised successfully.")
+        except Exception as exc:
+            logger.error("Failed to initialise OncoRAGRetriever: %s", exc)
+            raise
+    return _retriever_instance
+
+
+# ---------------------------------------------------------------------------
+# Document Grading (CRAG core)
+# ---------------------------------------------------------------------------
+
+def _grade_document(document_text: str, query: str, tier: int = 1) -> bool:
+    """Grade a single retrieved document for relevance.
+
+    Uses the Tier 1 (fast) model for binary classification.
+
+    Args:
+        document_text: The document text to evaluate.
+        query: The clinical query.
+        tier: Model tier to use for grading (default: 1 for speed).
+
+    Returns:
+        True if the document is RELEVANT, False otherwise.
+    """
+    # Rule-based shortcut: if the query contains a core cancer type and the document mentions it,
+    # we favor relevance to avoid model hallucination/rejection.
+    query_lower = query.lower()
+    doc_lower = document_text.lower()
+    
+    # Simple semantic overlap check: ignore generic terms
+    ignore_terms = {"treatment", "recommendation", "guidelines", "triage", "management", "clinical", "oncology"}
+    core_terms = [t for t in query_lower.split() if len(t) > 4 and t not in ignore_terms]
+    term_match = any(term in doc_lower for term in core_terms)
+
+    system_prompt = (
+        "You are an expert Oncology Clinical Analyst. "
+        "Your task is to evaluate if a retrieved medical document is RELEVANT to a patient's clinical query. "
+        "RELEVANCE CRITERIA:\n"
+        "1. The document discusses the specific cancer type or its precursors.\n"
+        "2. The document mentions treatment protocols, staging, or diagnostic criteria relevant to the case.\n"
+        "3. Synonyms are allowed (e.g., 'Uterine Cancer' is relevant to 'Endometrial Adenocarcinoma').\n\n"
+        "Output ONLY 'RELEVANT' or 'IRRELEVANT'. No explanation."
+    )
+    user_prompt = (
+        f"Patient Query/Context: {query}\n\n"
+        f"Document Snippet:\n--- START ---\n{document_text[:2000]}\n--- END ---\n\n"
+        f"Is this document relevant? (RELEVANT/IRRELEVANT):"
+    )
+
+    try:
+        response = call_tier_model(
+            tier=tier,
+            system_prompt=system_prompt,
+            user_prompt=user_prompt,
+            max_tokens=10,
+            temperature=0.0,
+        )
+        is_relevant = "RELEVANT" in response.upper()
+        logger.info("Doc Grade: %s (Term Match: %s) -> Query: %s...", "RELEVANT" if is_relevant else "IRRELEVANT", term_match, query[:30])
+        
+        # Boost: if model says IRRELEVANT but there is a strong term match, we might want to override.
+        # This ensures we don't drop guidelines that explicitly mention the cancer type.
+        if not is_relevant and term_match:
+            logger.debug("Model rejected doc but keyword match found. Overriding to RELEVANT for recall.")
+            return True
+            
+        return is_relevant
+    except Exception as exc:
+        logger.warning("Document grading failed: %s — defaulting to RELEVANT.", exc)
+        return True  # Fail open: include document if grading fails
+
+
+
+def _rewrite_query(
+    original_query: str,
+    entities: Dict[str, Any],
+    attempt: int,
+) -> str:
+    """Broaden the query for a retry attempt.
+
+    Uses deterministic broadening rather than LLM-based rewriting
+    for speed and predictability.
+
+    Args:
+        original_query: The query that yielded insufficient results.
+        entities: Extracted clinical entities.
+        attempt: The retry attempt number (1-indexed).
+
+    Returns:
+        A broadened query string.
+    """
+    cancer = entities.get("cancer_type", "Unknown")
+    stage = entities.get("stage", "Unknown")
+    mutations = entities.get("mutations", [])
+
+    if attempt == 1:
+        # Broadening strategy: remove stage specificity, keep cancer + mutations
+        parts = [cancer]
+        if mutations:
+            parts.append(f"mutations {' '.join(mutations)}")
+        parts.append("treatment guidelines evidence-based recommendations")
+        rewritten = " ".join(parts)
+        logger.info("Query rewrite attempt %d: %s → %s", attempt, original_query, rewritten)
+        return rewritten
+
+    # Attempt 2+: maximally broad
+    rewritten = f"{cancer} oncology clinical guidelines management"
+    logger.info("Query rewrite attempt %d (maximal broadening): %s", attempt, rewritten)
+    return rewritten
+
+
+# ---------------------------------------------------------------------------
+# Parallel Evidence Gathering
+# ---------------------------------------------------------------------------
+
+def _fetch_api_evidence(entities: Dict[str, Any]) -> Dict[str, List[str]]:
+    """Fetch genomic and clinical trial evidence in parallel.
+
+    Calls CIViC API and ClinicalTrials.gov concurrently for MI300X
+    throughput optimisation.
+
+    Args:
+        entities: Extracted clinical entities.
+
+    Returns:
+        Dict with "genomic_evidence" and "clinical_trials" lists.
+    """
+    results: Dict[str, List[str]] = {
+        "genomic_evidence": [],
+        "clinical_trials": [],
+    }
+
+    mutations = entities.get("mutations", [])
+    cancer = entities.get("cancer_type", "Unknown")
+
+    def fetch_civic():
+        """Fetch genomic evidence from CIViC."""
+        try:
+            from rag_engine.api_clients import CivicAPIClient
+            client = CivicAPIClient()
+            evidence = []
+            for mutation in mutations:
+                civic_results = client.search_variant_evidence(mutation, cancer)
+                for r in civic_results:
+                    evidence.append(
+                        f"[CIViC] {mutation}: {r.get('summary', 'No summary available')}"
+                    )
+            return evidence
+        except Exception as exc:
+            logger.warning("CIViC API failed: %s", exc)
+            return []
+
+    def fetch_trials():
+        """Fetch active clinical trials."""
+        try:
+            from rag_engine.api_clients import ClinicalTrialsClient
+            client = ClinicalTrialsClient()
+            trial_results = client.search_trials(cancer, mutations)
+            return [
+                f"[ClinicalTrials.gov] {t.get('title', 'Unknown')}: {t.get('status', '?')}"
+                for t in trial_results
+            ]
+        except Exception as exc:
+            logger.warning("ClinicalTrials.gov API failed: %s", exc)
+            return []
+
+    with ThreadPoolExecutor(max_workers=2) as executor:
+        futures = {
+            executor.submit(fetch_civic): "genomic_evidence",
+            executor.submit(fetch_trials): "clinical_trials",
+        }
+        for future in as_completed(futures):
+            key = futures[future]
+            try:
+                results[key] = future.result()
+            except Exception as exc:
+                logger.error("Parallel fetch error (%s): %s", key, exc)
+
+    return results
+
+
+# ---------------------------------------------------------------------------
+# Corrective RAG Node
+# ---------------------------------------------------------------------------
+
+# Minimum relevant documents required to proceed
+_MIN_RELEVANT_DOCS = 2
+# Maximum query rewrite attempts
+_MAX_REWRITES = 1
+
+
+def corrective_rag_node(state: AgentState) -> Dict[str, Any]:
+    """Execute the Corrective RAG pipeline.
+
+    Pipeline:
+      1. Build structured query from extracted entities.
+      2. Retrieve top-K candidates from ChromaDB.
+      3. Grade each document for relevance.
+      4. If insufficient relevant docs → rewrite query and retry.
+      5. Fetch API evidence in parallel (CIViC + ClinicalTrials).
+      6. Compute confidence metrics.
+
+    Args:
+        state: Current LangGraph state.
+
+    Returns:
+        State update with rag_context, sources, confidence, and metrics.
+    """
+    entities: Dict[str, Any] = state.get("extracted_entities", {})
+    clinical_text: str = state.get("clinical_text", "")
+    selected_tier: int = state.get("selected_tier", 1)
+
+    # --- Build initial query ---
+    cancer = entities.get("cancer_type", "Unknown")
+    stage = entities.get("stage", "Unknown")
+    mutations = ", ".join(entities.get("mutations", []))
+
+    query_parts = []
+    if cancer != "Unknown":
+        query_parts.append(cancer)
+    else:
+        # Fallback: use first 100 chars of clinical text for vector search
+        query_parts.append(clinical_text[:100].replace("\n", " "))
+    
+    if stage != "Unknown":
+        query_parts.append(stage)
+    if mutations:
+        query_parts.append(f"mutations: {mutations}")
+    query_parts.append("treatment recommendation guidelines triage")
+    query = " ".join(query_parts)
+
+    rewrite_count = 0
+    relevant_docs: List[Dict[str, Any]] = []
+
+    try:
+        retriever = _get_retriever()
+
+        # --- Retrieve + Grade loop ---
+        for attempt in range(1 + _MAX_REWRITES):
+            if attempt > 0:
+                query = _rewrite_query(query, entities, attempt)
+                rewrite_count += 1
+
+            # Retrieve candidates
+            raw_results = retriever.query(query, n_results=8)
+
+            # Grade documents in parallel for MI300X/API efficiency
+            from concurrent.futures import ThreadPoolExecutor
+            
+            def _grade_doc_wrapper(r):
+                doc_text = r.get("text", "")
+                is_relevant = _grade_document(doc_text, query, tier=1)
+                return r if is_relevant else None
+
+            with ThreadPoolExecutor(max_workers=8) as executor:
+                results = list(executor.map(_grade_doc_wrapper, raw_results))
+            
+            graded = [r for r in results if r is not None]
+
+            logger.info(
+                "CRAG attempt %d: %d/%d documents graded RELEVANT (Parallel).",
+                attempt + 1, len(graded), len(raw_results),
+            )
+
+            if len(graded) >= _MIN_RELEVANT_DOCS:
+                relevant_docs = graded
+                break
+
+        # --- Format results ---
+        context_strings = []
+        source_strings = []
+        for r in relevant_docs:
+            context_strings.append(
+                f"[Source: {r['source']}, Page: {r.get('page', '?')}, "
+                f"Section: {r.get('header', 'Unknown')}]\n{r['text']}"
+            )
+            source_strings.append(
+                f"- **{r['source']}** (Page {r.get('page', '?')}): "
+                f"{r.get('header', 'Unknown')}"
+            )
+
+        # --- Confidence metrics ---
+        ce_scores = [
+            r["cross_encoder_score"]
+            for r in relevant_docs
+            if "cross_encoder_score" in r
+        ]
+        mean_confidence = sum(ce_scores) / len(ce_scores) if ce_scores else 0.0
+
+    except Exception as exc:
+        logger.error("RAG retrieval failed: %s", exc)
+        context_strings = []
+        source_strings = []
+        relevant_docs = []
+        mean_confidence = 0.0
+        rewrite_count = 0
+
+    # --- Parallel API evidence ---
+    api_results = _fetch_api_evidence(entities)
+
+    return {
+        "rag_context": context_strings,
+        "rag_sources": source_strings,
+        "graph_rag_context": [],  # Future: knowledge graph integration
+        "api_evidence_context": (
+            api_results.get("genomic_evidence", [])
+            + api_results.get("clinical_trials", [])
+        ),
+        "rag_confidence": round(mean_confidence, 4),
+        "rag_retrieval_count": len(context_strings),
+        "rag_grading_pass_count": len(relevant_docs),
+        "rag_query_rewrites": rewrite_count,
+    }

agents/critic.py

@@ -0,0 +1,290 @@
+"""
+Critic Node — Reflexion-pattern validation for clinical recommendations.
+
+Design pattern: Reflexion (Shinn et al. 2023)
+  - Generator (Specialist) → Critic loop
+  - Critic evaluates: entailment, completeness, formatting
+  - If FAIL → specific feedback injected back to Specialist for retry
+  - Max 2 iterations before safe fallback
+
+Layer 1: Rule-based checks (deterministic, no LLM needed)
+Layer 2: LLM-based entailment verification (Tier 1 for speed)
+"""
+
+import logging
+import re
+from typing import Dict, Any, List
+
+from .state import AgentState
+from .tools import call_tier_model
+
+logger = logging.getLogger(__name__)
+
+
+# Maximum critic attempts before triggering safe fallback
+MAX_CRITIC_ATTEMPTS = 2
+
+# Required semantic concepts in a well-formed recommendation.
+# Each entry is a list of synonyms — at least ONE must appear.
+_REQUIRED_CONCEPTS = [
+    # Clinical findings / presentation
+    ["hallazgos", "findings", "presentación", "presentation", "clinical findings"],
+    # Diagnostic validation
+    ["diagnóstic", "diagnostic", "validación", "biopsia", "biopsy", "patholog", "patolog"],
+    # Management / treatment options
+    ["manejo", "management", "tratamiento", "treatment", "opciones", "options", "histerectom", "hysterectom", "surgery", "cirugía"],
+    # Final recommendation
+    ["recomendación", "recommendation", "conclusi", "next step"],
+]
+
+
+# ---------------------------------------------------------------------------
+# Layer 1: Deterministic checks (no LLM)
+# ---------------------------------------------------------------------------
+
+def _check_formatting(recommendation: str) -> tuple[bool, str]:
+    """Verify the recommendation contains required structural sections.
+
+    Uses flexible semantic matching instead of exact section headers,
+    so the model can use different header styles and still pass.
+
+    Args:
+        recommendation: The specialist's output text.
+
+    Returns:
+        Tuple of (passed, feedback_message).
+    """
+    text_lower = recommendation.lower()
+    missing_concepts = []
+
+    for synonyms in _REQUIRED_CONCEPTS:
+        if not any(syn in text_lower for syn in synonyms):
+            missing_concepts.append(synonyms[0])
+
+    if missing_concepts:
+        feedback = (
+            f"FORMATTING: Missing required concepts: {', '.join(missing_concepts)}. "
+            "Please include all sections: Hallazgos Clínicos, Validación Diagnóstica, "
+            "Análisis de Estadificación, Opciones de Manejo, Recomendación Final."
+        )
+        return False, feedback
+
+    return True, ""
+
+
+def _check_safety_phrases(recommendation: str) -> tuple[bool, str]:
+    """Check for known unsafe patterns (e.g., inventing dosages without sources).
+
+    Args:
+        recommendation: The specialist's output text.
+
+    Returns:
+        Tuple of (passed, feedback_message).
+    """
+    # Detect unsupported dosage patterns without source citations
+    dosage_pattern = re.compile(
+        r"\b\d+\s*(mg|mg/m2|mg/kg|mcg|IU|units)\b",
+        re.IGNORECASE,
+    )
+    dosages_found = dosage_pattern.findall(recommendation)
+
+    if dosages_found and "[source" not in recommendation.lower():
+        return False, (
+            "SAFETY: Specific dosages were mentioned without explicit source citations. "
+            "Either cite the guideline source for each dosage or remove the specific numbers."
+        )
+
+    return True, ""
+
+
+def _check_diagnostic_rigor(recommendation: str, clinical_text: str) -> tuple[bool, str]:
+    """Ensure no premature treatment is recommended without a confirmed diagnosis.
+
+    Args:
+        recommendation: The specialist's output text.
+        clinical_text: The original clinical input.
+
+    Returns:
+        Tuple of (passed, feedback_message).
+    """
+    text_lower = clinical_text.lower()
+    rec_lower = recommendation.lower()
+
+    # Detect if a biopsy/pathology was mentioned in the clinical text
+    pathology_keywords = [
+        "biopsia", "patología", "pathology", "biopsy", "histolog",
+        "legrado", "malign", "adenocarcinoma", "carcinoma", "sarcoma",
+        "linfoma", "lymphoma", "melanoma", "confirms", "confirma",
+        "diagnosed", "diagnosticado",
+    ]
+    has_pathology = any(word in text_lower for word in pathology_keywords)
+
+    # Treatment keywords
+    treatment_keywords = [
+        "cirugía", "radioterapia", "quimioterapia", "surgery",
+        "radiation", "chemotherapy", "histerectomía", "hysterectomy",
+    ]
+
+    if not has_pathology:
+        found_treatments = [kw for kw in treatment_keywords if kw in rec_lower]
+        if found_treatments:
+            feedback = (
+                f"DIAGNOSTIC RIGOR: Recommended treatments ({', '.join(found_treatments)}) "
+                "but no pathology/biopsy confirmation was found in the clinical text. "
+                "You MUST request a diagnostic procedure (e.g., biopsy) first."
+            )
+            return False, feedback
+
+    return True, ""
+
+
+# ---------------------------------------------------------------------------
+# Layer 2: LLM-based entailment check
+# ---------------------------------------------------------------------------
+
+def _check_entailment(
+    recommendation: str,
+    context: List[str],
+) -> tuple[bool, str]:
+    """Verify the recommendation is entailed by the RAG context.
+
+    Uses Tier 1 (fast model) for binary entailment classification.
+
+    Args:
+        recommendation: The specialist's output text.
+        context: The RAG context strings.
+
+    Returns:
+        Tuple of (passed, feedback_message).
+    """
+    context_summary = "\n---\n".join(context)
+
+    system_prompt = (
+        "You are a clinical safety auditor. Verify if a treatment recommendation "
+        "is STRICTLY grounded in the provided clinical guidelines context.\n\n"
+        "Check for:\n"
+        "1. Any drug, treatment, or procedure mentioned that is NOT in the context.\n"
+        "2. Any dosage or protocol that contradicts the context.\n"
+        "3. Any claim presented as fact that lacks support in the context.\n\n"
+        "Output a JSON object with two keys:\n"
+        '- "verdict": "PASS" or "FAIL"\n'
+        '- "issues": a list of specific issues found (empty list if PASS)\n\n'
+        "Output ONLY the JSON, nothing else."
+    )
+
+    user_prompt = (
+        f"Context:\n{context_summary}\n\n"
+        f"Recommendation:\n{recommendation}\n\n"
+        "Evaluate the recommendation against the context:"
+    )
+
+    try:
+        response = call_tier_model(
+            tier=1,
+            system_prompt=system_prompt,
+            user_prompt=user_prompt,
+            max_tokens=200,
+            temperature=0.0,
+        )
+
+        response_upper = response.upper()
+        if "FAIL" in response_upper:
+            feedback = f"ENTAILMENT: {response}"
+            return False, feedback
+
+        return True, ""
+
+    except Exception as exc:
+        logger.warning("Entailment check failed: %s — defaulting to PASS.", exc)
+        return True, ""  # Fail open if entailment check itself fails
+
+
+# ---------------------------------------------------------------------------
+# Critic Node
+# ---------------------------------------------------------------------------
+
+def critic_node(state: AgentState) -> Dict[str, Any]:
+    """Validate the specialist's recommendation for safety and completeness.
+
+    Runs four layers of checks:
+      1. Formatting (deterministic — flexible concept matching)
+      2. Safety phrases (deterministic — dosage citation check)
+      3. Diagnostic rigor (deterministic — no treatment without pathology)
+      4. Entailment (LLM-based — only if layers 1-3 pass)
+
+    If any check fails, returns FAIL with specific feedback.
+    The graph will loop back to the specialist if attempts < max.
+
+    Args:
+        state: Current LangGraph state.
+
+    Returns:
+        State update with critic_verdict, critic_feedback, critic_attempts.
+    """
+    recommendation = state.get("clinical_recommendation", "")
+    context = state.get("rag_context", [])
+    current_attempts = state.get("critic_attempts", 0)
+
+    # Track this attempt
+    new_attempts = current_attempts + 1
+
+    # Guard: empty recommendation
+    if not recommendation or not recommendation.strip():
+        logger.warning("Critic received empty recommendation — auto-FAIL.")
+        return {
+            "critic_verdict": "FAIL",
+            "critic_feedback": "SYSTEM: Specialist returned empty recommendation.",
+            "critic_attempts": new_attempts,
+        }
+
+    # Guard: recommendation is already the safe fallback phrase
+    if "información no concluyente" in recommendation.lower():
+        return {
+            "critic_verdict": "PASS",
+            "critic_feedback": "",
+            "critic_attempts": new_attempts,
+        }
+
+    # Guard: inference error
+    if "error en el sistema de inferencia" in recommendation.lower():
+        return {
+            "critic_verdict": "FAIL",
+            "critic_feedback": "SYSTEM: Inference engine error — cannot validate.",
+            "critic_attempts": new_attempts,
+        }
+
+    # --- Layer 1: Formatting check ---
+    fmt_pass, fmt_feedback = _check_formatting(recommendation)
+
+    # --- Layer 2: Safety check ---
+    safety_pass, safety_feedback = _check_safety_phrases(recommendation)
+
+    # --- Layer 3: Diagnostic Rigor check ---
+    clinical_text = state.get("clinical_text", "")
+    rigor_pass, rigor_feedback = _check_diagnostic_rigor(recommendation, clinical_text)
+
+    # --- Layer 4: Entailment check (only if layers 1-3 pass) ---
+    entailment_pass = True
+    entailment_feedback = ""
+    if fmt_pass and safety_pass and rigor_pass and context:
+        entailment_pass, entailment_feedback = _check_entailment(
+            recommendation, context
+        )
+
+    # --- Aggregate verdict ---
+    all_passed = fmt_pass and safety_pass and rigor_pass and entailment_pass
+    feedbacks = [f for f in [fmt_feedback, safety_feedback, rigor_feedback, entailment_feedback] if f]
+
+    verdict = "PASS" if all_passed else "FAIL"
+    combined_feedback = "\n".join(feedbacks)
+
+    logger.info(
+        "Critic verdict: %s (attempt %d/%d). Issues: %d",
+        verdict, new_attempts, MAX_CRITIC_ATTEMPTS, len(feedbacks),
+    )
+
+    return {
+        "critic_verdict": verdict,
+        "critic_feedback": combined_feedback,
+        "critic_attempts": new_attempts,
+    }

agents/formatter.py

@@ -0,0 +1,182 @@
+"""
+Formatter & Fallback Nodes — Structured output and safe degradation.
+
+Formatter: transforms the validated recommendation into a structured
+format optimised for the Gradio UI, including confidence reports and
+source citations.
+
+Fallback: safe degradation when RAG or reasoning fails, following the
+Anti-Hallucination Policy (Rule #39).
+"""
+
+import logging
+from datetime import datetime, timezone
+from typing import Dict, Any, List
+
+from .state import AgentState
+from .tools import get_tier_spec
+
+logger = logging.getLogger(__name__)
+
+
+# ---------------------------------------------------------------------------
+# Response Formatter Node
+# ---------------------------------------------------------------------------
+
+def formatter_node(state: AgentState) -> Dict[str, Any]:
+    """Transform the validated recommendation into structured UI output.
+
+    Produces:
+      - formatted_recommendation: Markdown with metadata header
+      - confidence_report: Dict of all quality metrics
+      - source_citations: Formatted bibliography
+
+    Args:
+        state: Current LangGraph state.
+
+    Returns:
+        State update with formatted output, confidence report, and citations.
+    """
+    recommendation = state.get("clinical_recommendation", "")
+    tier = state.get("selected_tier", 1)
+    spec = get_tier_spec(tier)
+    rag_confidence = state.get("rag_confidence", 0.0)
+    critic_attempts = state.get("critic_attempts", 0)
+    complexity_score = state.get("complexity_score", 0.0)
+    rag_sources = state.get("rag_sources", [])
+    rag_count = state.get("rag_retrieval_count", 0)
+    rag_graded = state.get("rag_grading_pass_count", 0)
+    rag_rewrites = state.get("rag_query_rewrites", 0)
+    api_evidence = state.get("api_evidence_context", [])
+    entities = state.get("extracted_entities", {})
+
+    # --- Confidence report ---
+    confidence_report: Dict[str, Any] = {
+        "tier_used": tier,
+        "tier_name": spec.name,
+        "model_id": spec.model_id,
+        "complexity_score": complexity_score,
+        "rag_confidence": rag_confidence,
+        "rag_retrieval_count": rag_count,
+        "rag_graded_relevant": rag_graded,
+        "rag_query_rewrites": rag_rewrites,
+        "critic_iterations": critic_attempts,
+        "api_evidence_count": len(api_evidence),
+        "timestamp": datetime.now(timezone.utc).isoformat(),
+    }
+
+    # --- Confidence level label ---
+    if rag_confidence >= 0.7:
+        confidence_label = "🟢 Alta"
+    elif rag_confidence >= 0.4:
+        confidence_label = "🟡 Media"
+    else:
+        confidence_label = "🔴 Baja"
+
+    # --- Formatted recommendation with metadata header ---
+    header = (
+        f"---\n"
+        f"**OncoAgent — Recomendación Clínica**\n"
+        f"📊 Modelo: {spec.name} (Tier {tier}) | "
+        f"Confianza RAG: {confidence_label} ({rag_confidence:.2f}) | "
+        f"Iteraciones Críticas: {critic_attempts}\n"
+        f"🧬 Tipo: {entities.get('cancer_type', 'N/A')} | "
+        f"Estadío: {entities.get('stage', 'N/A')} | "
+        f"Mutaciones: {', '.join(entities.get('mutations', [])) or 'N/A'}\n"
+        f"---\n\n"
+    )
+
+    formatted = header + recommendation
+
+    # --- Source citations ---
+    citations = []
+    if rag_sources:
+        citations.append("### Fuentes Clínicas (RAG)")
+        citations.extend(rag_sources)
+
+    if api_evidence:
+        citations.append("\n### Evidencia Adicional (APIs)")
+        citations.extend([f"- {e}" for e in api_evidence])
+
+    # --- Safety status ---
+    safety_status = "Validated against clinical oncology guidelines"
+
+    return {
+        "formatted_recommendation": formatted,
+        "confidence_report": confidence_report,
+        "source_citations": citations,
+        "safety_status": safety_status,
+        "is_safe": True,
+    }
+
+
+# ---------------------------------------------------------------------------
+# Fallback Node (Safe Degradation)
+# ---------------------------------------------------------------------------
+
+_SAFE_MESSAGE = (
+    "---\n"
+    "**OncoAgent — Resultado No Concluyente**\n"
+    "---\n\n"
+    "## ⚠️ Información no concluyente en las guías provistas.\n\n"
+    "El sistema no pudo generar una recomendación clínica confiable "
+    "para este caso por una de las siguientes razones:\n\n"
+    "1. No se encontró evidencia suficiente en las guías clínicas cargadas.\n"
+    "2. La recomendación generada no pasó la validación de seguridad.\n"
+    "3. El caso requiere revisión clínica especializada fuera del alcance "
+    "de las guías disponibles.\n\n"
+    "**Acción recomendada:** Consulte con un oncólogo especialista para "
+    "una evaluación personalizada.\n"
+)
+
+
+def fallback_node(state: AgentState) -> Dict[str, Any]:
+    """Generate a safe fallback response when the pipeline cannot produce
+    a reliable recommendation.
+
+    This node is triggered when:
+      - RAG retrieval yields insufficient relevant documents
+      - The critic fails after max iterations
+      - The input is too short or unintelligible
+
+    Args:
+        state: Current LangGraph state.
+
+    Returns:
+        State update with safe fallback response and diagnostic info.
+    """
+    # Determine why we fell back
+    routing = state.get("routing_decision", "")
+    rag_count = state.get("rag_retrieval_count", 0)
+    critic_verdict = state.get("critic_verdict", "")
+    critic_attempts = state.get("critic_attempts", 0)
+
+    reasons = []
+    if routing == "insufficient":
+        reasons.append("Input too short or unintelligible for clinical triage.")
+    if rag_count == 0:
+        reasons.append("No relevant documents found in clinical guidelines database.")
+    if critic_verdict == "FAIL" and critic_attempts >= 2:
+        reasons.append(
+            f"Recommendation failed safety validation after {critic_attempts} attempts."
+        )
+    if not reasons:
+        reasons.append("Unknown system error — safe fallback triggered.")
+
+    fallback_reason = " | ".join(reasons)
+    logger.warning("Fallback triggered: %s", fallback_reason)
+
+    return {
+        "formatted_recommendation": _SAFE_MESSAGE,
+        "clinical_recommendation": "Información no concluyente en las guías provistas.",
+        "confidence_report": {
+            "tier_used": state.get("selected_tier", 0),
+            "fallback": True,
+            "reason": fallback_reason,
+            "timestamp": datetime.now(timezone.utc).isoformat(),
+        },
+        "source_citations": [],
+        "fallback_reason": fallback_reason,
+        "safety_status": f"Fallback: {fallback_reason}",
+        "is_safe": False,
+    }

agents/graph.py

@@ -0,0 +1,260 @@
+"""
+OncoAgent LangGraph — SOTA Multi-Agent Orchestration Graph.
+
+Architecture synthesised from:
+  - Claude Code: deterministic harness + sub-agent delegation
+  - Hermes Agent: structured tool calling + persistent state
+  - Corrective RAG: graded retrieval with query rewriting
+  - Reflexion: generator ↔ critic loop with max iterations
+  - Model Tiering: Qwen3.5-9B (fast) ↔ Qwen3.6-27B (deep reasoning)
+
+Topology:
+  Router → Ingestion → Corrective RAG → Specialist ↔ Critic → HITL Gate → Formatter
+                                                                    ↓
+                                                                Fallback
+
+Conditional edges:
+  - Router: routes "insufficient" directly to fallback
+  - CRAG: routes insufficient docs to fallback
+  - Critic: loops back to specialist (max 2) or to fallback
+  - HITL: routes high-acuity to interrupt, others to formatter
+"""
+
+import logging
+from langgraph.graph import StateGraph, END
+from langgraph.checkpoint.memory import MemorySaver
+
+from .state import AgentState
+from .router import router_node
+from .nodes import data_ingestion_node
+from .corrective_rag import corrective_rag_node
+from .specialist import specialist_node
+from .critic import critic_node, MAX_CRITIC_ATTEMPTS
+from .formatter import formatter_node, fallback_node
+
+logger = logging.getLogger(__name__)
+
+
+# ---------------------------------------------------------------------------
+# Conditional edge functions
+# ---------------------------------------------------------------------------
+
+def _route_after_router(state: AgentState) -> str:
+    """Route based on the router's complexity classification.
+
+    Returns:
+        Node name to transition to.
+    """
+    decision = state.get("routing_decision", "simple")
+    if decision == "insufficient":
+        logger.info("Router → Fallback (insufficient input)")
+        return "fallback"
+    # Both "simple" and "complex" proceed to ingestion
+    return "ingestion"
+
+
+def _route_after_crag(state: AgentState) -> str:
+    """Route based on CRAG retrieval results.
+
+    If insufficient relevant documents were found (even after rewrites),
+    route directly to fallback.
+
+    Returns:
+        Node name to transition to.
+    """
+    graded_count = state.get("rag_grading_pass_count", 0)
+    retrieval_count = state.get("rag_retrieval_count", 0)
+
+    if retrieval_count == 0 and graded_count == 0:
+        logger.info("CRAG → Fallback (no relevant documents)")
+        return "fallback"
+
+    return "specialist"
+
+
+def _route_after_critic(state: AgentState) -> str:
+    """Route based on the critic's verdict and attempt count.
+
+    - PASS → proceed to HITL gate
+    - FAIL + attempts < max → loop back to specialist
+    - FAIL + attempts >= max → fallback
+
+    Returns:
+        Node name to transition to.
+    """
+    verdict = state.get("critic_verdict", "FAIL")
+    attempts = state.get("critic_attempts", 0)
+
+    if verdict == "PASS":
+        logger.info("Critic → HITL Gate (PASS on attempt %d)", attempts)
+        return "hitl_gate"
+
+    if attempts >= MAX_CRITIC_ATTEMPTS:
+        logger.warning(
+            "Critic → Fallback (FAIL after %d/%d attempts)",
+            attempts, MAX_CRITIC_ATTEMPTS,
+        )
+        return "fallback"
+
+    logger.info(
+        "Critic → Specialist retry (FAIL, attempt %d/%d)",
+        attempts, MAX_CRITIC_ATTEMPTS,
+    )
+    return "specialist"
+
+
+def _route_after_hitl(state: AgentState) -> str:
+    """Route based on acuity level and HITL requirements.
+
+    For the hackathon, high-acuity cases are flagged but auto-proceed.
+    In production, this would use LangGraph's interrupt() for real
+    clinician approval.
+
+    Returns:
+        Node name to transition to.
+    """
+    # For now, always proceed to formatter
+    # In production: if hitl_required and not hitl_approved → interrupt
+    return "formatter"
+
+
+# ---------------------------------------------------------------------------
+# HITL Gate Node
+# ---------------------------------------------------------------------------
+
+def hitl_gate_node(state: AgentState) -> dict:
+    """Determine if the case requires Human-in-the-Loop approval.
+
+    Acuity classification:
+      - high: Stage IV + rare mutations → requires clinician review
+      - medium: Stage III or complex → flagged but auto-proceeds
+      - low: Standard cases → auto-proceeds
+
+    Args:
+        state: Current LangGraph state.
+
+    Returns:
+        State update with acuity_level, hitl_required, hitl_approved.
+    """
+    entities = state.get("extracted_entities", {})
+    complexity = state.get("complexity_score", 0.0)
+    stage = entities.get("stage", "Unknown").upper()
+
+    # Determine acuity
+    if "IV" in stage and complexity >= 0.6:
+        acuity = "high"
+        hitl_required = True
+    elif "III" in stage or complexity >= 0.4:
+        acuity = "medium"
+        hitl_required = False
+    else:
+        acuity = "low"
+        hitl_required = False
+
+    logger.info(
+        "HITL Gate: acuity=%s, hitl_required=%s, complexity=%.2f",
+        acuity, hitl_required, complexity,
+    )
+
+    return {
+        "acuity_level": acuity,
+        "hitl_required": hitl_required,
+        "hitl_approved": not hitl_required,  # Auto-approve non-HITL cases
+    }
+
+
+# ---------------------------------------------------------------------------
+# Graph Builder
+# ---------------------------------------------------------------------------
+
+def build_oncoagent_graph() -> StateGraph:
+    """Build the SOTA OncoAgent LangGraph state machine.
+
+    Topology:
+      START → router → (ingestion | fallback)
+                          ↓
+                    corrective_rag → (specialist | fallback)
+                          ↓
+                    specialist ↔ critic (max 2 loops)
+                          ↓
+                    hitl_gate → formatter → END
+                          ↓
+                    fallback → END
+
+    Returns:
+        Compiled LangGraph state machine.
+    """
+    workflow = StateGraph(AgentState)
+
+    # --- Define Nodes ---
+    workflow.add_node("router", router_node)
+    workflow.add_node("ingestion", data_ingestion_node)
+    workflow.add_node("corrective_rag", corrective_rag_node)
+    workflow.add_node("specialist", specialist_node)
+    workflow.add_node("critic", critic_node)
+    workflow.add_node("hitl_gate", hitl_gate_node)
+    workflow.add_node("formatter", formatter_node)
+    workflow.add_node("fallback", fallback_node)
+
+    # --- Define Edges ---
+    # Entry point
+    workflow.set_entry_point("router")
+
+    # Router → Ingestion or Fallback (conditional)
+    workflow.add_conditional_edges(
+        "router",
+        _route_after_router,
+        {
+            "ingestion": "ingestion",
+            "fallback": "fallback",
+        },
+    )
+
+    # Ingestion → Corrective RAG (always)
+    workflow.add_edge("ingestion", "corrective_rag")
+
+    # Corrective RAG → Specialist or Fallback (conditional)
+    workflow.add_conditional_edges(
+        "corrective_rag",
+        _route_after_crag,
+        {
+            "specialist": "specialist",
+            "fallback": "fallback",
+        },
+    )
+
+    # Specialist → Critic (always)
+    workflow.add_edge("specialist", "critic")
+
+    # Critic → HITL Gate, Specialist (retry), or Fallback (conditional)
+    workflow.add_conditional_edges(
+        "critic",
+        _route_after_critic,
+        {
+            "hitl_gate": "hitl_gate",
+            "specialist": "specialist",
+            "fallback": "fallback",
+        },
+    )
+
+    # HITL Gate → Formatter (conditional, future: interrupt for clinician)
+    workflow.add_conditional_edges(
+        "hitl_gate",
+        _route_after_hitl,
+        {
+            "formatter": "formatter",
+        },
+    )
+
+    # Terminal edges
+    workflow.add_edge("formatter", END)
+    workflow.add_edge("fallback", END)
+
+    # Compile with recursion limit (Rule #20: strict limit for loops)
+    memory = MemorySaver()
+    compiled = workflow.compile(
+        checkpointer=memory,
+    )
+
+    logger.info("OncoAgent graph compiled successfully (8 nodes, SOTA topology).")
+    return compiled

agents/memory.py

@@ -0,0 +1,162 @@
+"""
+Per-patient session memory for OncoAgent.
+
+Design inspired by Hermes Agent's persistent memory:
+  - Each patient gets an isolated profile with their own clinical history.
+  - Memory is scoped per ``patient_id``, never global.
+  - Thread-safe via a simple dict-based store (swap for Redis/SQLite
+    in production if needed).
+
+Usage:
+    store = PatientMemoryStore()
+    store.save_interaction(patient_id="P001", interaction={...})
+    history = store.get_history(patient_id="P001")
+"""
+
+import logging
+import uuid
+from datetime import datetime, timezone
+from typing import Dict, Any, List, Optional
+from dataclasses import dataclass, field
+
+logger = logging.getLogger(__name__)
+
+
+@dataclass
+class PatientProfile:
+    """Isolated memory profile for a single patient.
+
+    Attributes:
+        patient_id: Unique identifier for the patient.
+        created_at: ISO timestamp of profile creation.
+        interactions: Ordered list of past query/response pairs.
+        metadata: Arbitrary metadata (e.g., preferred language).
+    """
+
+    patient_id: str
+    created_at: str = field(
+        default_factory=lambda: datetime.now(timezone.utc).isoformat()
+    )
+    interactions: List[Dict[str, Any]] = field(default_factory=list)
+    metadata: Dict[str, Any] = field(default_factory=dict)
+
+    def add_interaction(self, interaction: Dict[str, Any]) -> None:
+        """Append an interaction to the patient's history.
+
+        Args:
+            interaction: Dict with at minimum ``query`` and ``response`` keys.
+        """
+        interaction["timestamp"] = datetime.now(timezone.utc).isoformat()
+        interaction["interaction_id"] = str(uuid.uuid4())[:8]
+        self.interactions.append(interaction)
+        logger.debug(
+            "Patient %s: stored interaction #%d",
+            self.patient_id,
+            len(self.interactions),
+        )
+
+    def get_recent_context(self, n: int = 3) -> List[Dict[str, Any]]:
+        """Return the last *n* interactions for context injection.
+
+        Args:
+            n: Number of recent interactions to return.
+
+        Returns:
+            List of the most recent interactions (newest last).
+        """
+        return self.interactions[-n:]
+
+    def summary(self) -> str:
+        """Return a brief summary string for logging/UI display."""
+        return (
+            f"Patient {self.patient_id} | "
+            f"{len(self.interactions)} interactions | "
+            f"Created: {self.created_at}"
+        )
+
+
+class PatientMemoryStore:
+    """In-memory store for per-patient profiles.
+
+    For hackathon scope this uses a simple dict.  In production,
+    replace with SQLite / Redis for persistence across restarts.
+    """
+
+    def __init__(self) -> None:
+        self._profiles: Dict[str, PatientProfile] = {}
+
+    def get_or_create_profile(
+        self,
+        patient_id: Optional[str] = None,
+    ) -> PatientProfile:
+        """Retrieve an existing profile or create a new one.
+
+        Args:
+            patient_id: Existing patient ID. If None, generates a new one.
+
+        Returns:
+            The corresponding PatientProfile.
+        """
+        if patient_id is None:
+            patient_id = f"P-{str(uuid.uuid4())[:8].upper()}"
+
+        if patient_id not in self._profiles:
+            self._profiles[patient_id] = PatientProfile(patient_id=patient_id)
+            logger.info("Created new patient profile: %s", patient_id)
+
+        return self._profiles[patient_id]
+
+    def save_interaction(
+        self,
+        patient_id: str,
+        interaction: Dict[str, Any],
+    ) -> None:
+        """Save an interaction to a patient's profile.
+
+        Args:
+            patient_id: Target patient ID.
+            interaction: Dict with query/response data.
+        """
+        profile = self.get_or_create_profile(patient_id)
+        profile.add_interaction(interaction)
+
+    def get_history(
+        self,
+        patient_id: str,
+        n: Optional[int] = None,
+    ) -> List[Dict[str, Any]]:
+        """Retrieve a patient's interaction history.
+
+        Args:
+            patient_id: Target patient ID.
+            n: If provided, return only the last *n* interactions.
+
+        Returns:
+            List of interaction dicts.
+        """
+        profile = self._profiles.get(patient_id)
+        if profile is None:
+            return []
+        if n is not None:
+            return profile.get_recent_context(n)
+        return profile.interactions
+
+    def list_patients(self) -> List[str]:
+        """Return all known patient IDs."""
+        return list(self._profiles.keys())
+
+    def patient_count(self) -> int:
+        """Return the number of tracked patients."""
+        return len(self._profiles)
+
+
+# Module-level singleton
+_global_memory_store: Optional[PatientMemoryStore] = None
+
+
+def get_memory_store() -> PatientMemoryStore:
+    """Return the global PatientMemoryStore singleton."""
+    global _global_memory_store
+    if _global_memory_store is None:
+        _global_memory_store = PatientMemoryStore()
+    return _global_memory_store

agents/nodes.py

@@ -0,0 +1,195 @@
+"""
+LangGraph Node implementations for OncoAgent.
+
+This module retains the data ingestion node (PHI cleaning + entity extraction)
+and re-exports all other nodes from their dedicated modules for backward
+compatibility.
+
+Module organisation (SOTA redesign):
+  - agents/router.py       → Router Node (complexity classification)
+  - agents/corrective_rag.py → Corrective RAG Node (graded retrieval)
+  - agents/specialist.py   → Specialist Node (tier-adaptive reasoning)
+  - agents/critic.py       → Critic Node (reflexion validation)
+  - agents/formatter.py    → Formatter + Fallback Nodes
+  - agents/tools.py        → Shared vLLM client + tier calling
+  - agents/memory.py       → Per-patient session memory
+"""
+
+from typing import Dict, Any
+
+import re
+import logging
+
+from .state import AgentState
+
+logger = logging.getLogger(__name__)
+
+
+# ---------------------------------------------------------------------------
+# PHI Patterns (Zero-PHI Policy — Rule #39)
+# ---------------------------------------------------------------------------
+
+_PHI_PATTERNS = [
+    re.compile(r"\b\d{3}-\d{2}-\d{4}\b"),                 # SSN
+    re.compile(r"\b\d{2}/\d{2}/\d{4}\b"),                 # Date of birth
+    re.compile(r"[A-Za-z0-9._%+-]+@[A-Za-z0-9.-]+\.[A-Z|a-z]{2,}"),  # Email
+]
+
+
+# ---------------------------------------------------------------------------
+# Node 1: Data Ingestion — PHI cleaning & entity extraction
+# ---------------------------------------------------------------------------
+
+def data_ingestion_node(state: AgentState) -> Dict[str, Any]:
+    """Clean the input clinical text (Zero-PHI policy) and extract
+    key medical entities via rule-based heuristics.
+
+    Enhanced extraction includes:
+      - Cancer type identification (20+ types)
+      - TNM staging parsing
+      - Biomarker/mutation detection (15+ markers)
+      - Performance status detection (ECOG)
+      - Urgency signals
+
+    Args:
+        state: Current LangGraph state with ``clinical_text``.
+
+    Returns:
+        State update with ``extracted_entities`` and ``phi_detected``.
+    """
+    text: str = state.get("clinical_text", "")
+
+    # --- Zero-PHI check and redaction ---
+    phi_found = False
+    cleaned_text = text
+    for pattern in _PHI_PATTERNS:
+        if pattern.search(text):
+            phi_found = True
+            # Redact detected PHI
+            cleaned_text = pattern.sub("[REDACTED]", cleaned_text)
+    
+    if phi_found:
+        logger.warning("PHI detected and redacted from clinical input.")
+
+    # Use cleaned text for downstream processing
+    text = cleaned_text
+
+    # --- Rule-based entity extraction ---
+    extracted: Dict[str, Any] = {
+        "cancer_type": "Unknown",
+        "stage": "Unknown",
+        "mutations": [],
+        "ecog_status": "Unknown",
+        "urgency": "routine",
+    }
+
+    text_lower = text.lower()
+
+    # Cancer type heuristic (Explicit + Symptom-based risk)
+    cancer_keywords = {
+        "breast": "Breast Cancer",
+        "lung": "Lung Cancer",
+        "non-small cell": "Non-Small Cell Lung Cancer",
+        "small cell lung": "Small Cell Lung Cancer",
+        "colon": "Colon Cancer",
+        "colorectal": "Colorectal Cancer",
+        "prostate": "Prostate Cancer",
+        "pancreatic": "Pancreatic Cancer",
+        "hepatocellular": "Hepatocellular Carcinoma",
+        "hcc": "Hepatocellular Carcinoma",
+        "melanoma": "Melanoma",
+        "renal": "Renal Cell Carcinoma",
+        "bladder": "Bladder Cancer",
+        "ovarian": "Ovarian Cancer",
+        "cervical": "Cervical Cancer",
+        "thyroid": "Thyroid Cancer",
+        "leukemia": "Leukemia",
+        "lymphoma": "Lymphoma",
+        "myeloma": "Multiple Myeloma",
+        "sarcoma": "Sarcoma",
+        "glioma": "Glioma",
+        "glioblastoma": "Glioblastoma",
+        "esophageal": "Esophageal Cancer",
+        "gastric": "Gastric Cancer",
+        "cholangiocarcinoma": "Cholangiocarcinoma",
+        "mesothelioma": "Mesothelioma",
+        "uterine": "Uterine Cancer",
+        "endometrial": "Uterine Cancer",
+        # Symptom-based risk mapping (Triage mode) - Multilingual support
+        "menstru": "Uterine Cancer",
+        "vaginal": "Uterine Cancer",
+        "bleeding": "Uterine Cancer",
+        "sangrado": "Uterine Cancer",
+        "periods": "Uterine Cancer",
+        "periodo": "Uterine Cancer",
+        "postmenopausal": "Uterine Cancer",
+        "postmenopau": "Uterine Cancer",
+        "hemorragia": "Uterine Cancer",
+    }
+    for keyword, label in cancer_keywords.items():
+        if keyword in text_lower:
+            extracted["cancer_type"] = label
+            break
+
+    # Stage heuristic (supports TNM and simple staging)
+    stage_match = re.search(
+        r"stage\s+(I{1,3}V?|[1-4]|iv|iii|ii|i)\b",
+        text,
+        re.IGNORECASE,
+    )
+    if stage_match:
+        extracted["stage"] = f"Stage {stage_match.group(1).upper()}"
+
+    # TNM staging
+    tnm_match = re.search(
+        r"\b(T[0-4x]N[0-3x]M[01x])\b",
+        text,
+        re.IGNORECASE,
+    )
+    if tnm_match:
+        extracted["tnm"] = tnm_match.group(1).upper()
+
+    # Mutation heuristic (expanded)
+    mutations_found = re.findall(
+        r"\b(EGFR|ALK|KRAS|BRAF|HER2|TP53|BRCA[12]|PD-?L1|ROS1|MET|RET|"
+        r"NTRK|PIK3CA|MSI-?H|dMMR|FGFR[1-4]?|IDH[12]?|ERBB2|CDK[46]|"
+        r"PTEN|APC|VEGF|mTOR)\b",
+        text,
+        re.IGNORECASE,
+    )
+    if mutations_found:
+        extracted["mutations"] = list(set(m.upper() for m in mutations_found))
+
+    # ECOG Performance Status
+    ecog_match = re.search(
+        r"(?:ECOG|performance\s+status)\s*(?:of\s*)?(\d)",
+        text,
+        re.IGNORECASE,
+    )
+    if ecog_match:
+        extracted["ecog_status"] = f"ECOG {ecog_match.group(1)}"
+
+    # Urgency detection
+    urgency_keywords = [
+        "urgent", "emergency", "critical", "immediate",
+        "rapidly progressing", "acute", "life-threatening",
+    ]
+    for kw in urgency_keywords:
+        if kw in text_lower:
+            extracted["urgency"] = "urgent"
+            break
+
+    return {
+        "clinical_text": cleaned_text,
+        "extracted_entities": extracted,
+        "phi_detected": phi_found,
+    }
+
+
+# ---------------------------------------------------------------------------
+# Re-exports for backward compatibility
+# ---------------------------------------------------------------------------
+
+from .corrective_rag import corrective_rag_node as rag_retrieval_node  # noqa: E402, F401
+from .specialist import specialist_node as clinical_specialist_node  # noqa: E402, F401
+from .critic import critic_node as safety_validator_node  # noqa: E402, F401

agents/router.py

@@ -0,0 +1,161 @@
+"""
+Router Node — Complexity classification and model tier selection.
+
+Design pattern: Supervisor Routing (LangGraph SOTA)
+Inspired by:
+  - Claude Code: deterministic routing via structured logic, not free text
+  - Hermes Agent: structured JSON output for decisions
+
+The router classifies each clinical case into one of three categories:
+  - ``simple``:  Well-known cancer + standard staging → Tier 1 (9B)
+  - ``complex``: Rare cancer / multi-mutation / ambiguous staging → Tier 2 (27B)
+  - ``insufficient``: Input too short or unintelligible → direct fallback
+
+Supports manual tier override from the UI (user can force Tier 1 or 2).
+"""
+
+import logging
+import json
+from typing import Dict, Any, Optional
+
+from .state import AgentState
+
+logger = logging.getLogger(__name__)
+
+
+# ---------------------------------------------------------------------------
+# Complexity heuristics
+# ---------------------------------------------------------------------------
+
+# Cancer types considered "well-documented" with standard NCCN guidelines
+_COMMON_CANCERS = frozenset({
+    "breast cancer", "lung cancer", "colon cancer", "colorectal cancer",
+    "prostate cancer", "melanoma", "bladder cancer", "thyroid cancer",
+    "cervical cancer", "ovarian cancer", "gastric cancer",
+})
+
+# Cancer types considered rare or requiring deeper reasoning
+_RARE_CANCERS = frozenset({
+    "pancreatic cancer", "hepatocellular carcinoma", "sarcoma",
+    "glioma", "glioblastoma", "multiple myeloma", "renal cell carcinoma",
+    "esophageal cancer", "cholangiocarcinoma", "mesothelioma",
+    "neuroendocrine tumor", "adrenocortical carcinoma",
+})
+
+# Mutations that indicate multi-pathway complexity
+_COMPLEX_MUTATIONS = frozenset({
+    "EGFR", "ALK", "KRAS", "NTRK", "RET", "MET", "ROS1",
+    "PIK3CA", "MSI-H", "DMMR", "BRAF V600E",
+})
+
+# Minimum character count for a clinically meaningful input
+_MIN_INPUT_LENGTH = 30
+
+
+def _classify_complexity(
+    clinical_text: str,
+    entities: Dict[str, Any],
+) -> tuple[str, float, int]:
+    """Classify case complexity using rule-based heuristics.
+
+    Args:
+        clinical_text: Raw clinical text.
+        entities: Extracted entities from the ingestion node.
+
+    Returns:
+        Tuple of (routing_decision, complexity_score, recommended_tier).
+    """
+    # Gate: insufficient input
+    if len(clinical_text.strip()) < _MIN_INPUT_LENGTH:
+        logger.info("Input too short (%d chars) — routing to insufficient.", len(clinical_text))
+        return "insufficient", 0.0, 1
+
+    score = 0.0
+    cancer_type = entities.get("cancer_type", "Unknown").lower()
+    stage = entities.get("stage", "Unknown")
+    mutations = entities.get("mutations", [])
+
+    # --- Cancer type scoring ---
+    if cancer_type in _RARE_CANCERS:
+        score += 0.4
+    elif cancer_type == "unknown":
+        score += 0.3  # Unidentified cancer is inherently complex
+    # Common cancers add no complexity
+
+    # --- Stage scoring ---
+    if "IV" in stage.upper():
+        score += 0.25
+    elif "III" in stage.upper():
+        score += 0.15
+
+    # --- Mutation complexity ---
+    complex_muts = [m for m in mutations if m.upper() in _COMPLEX_MUTATIONS]
+    if len(complex_muts) >= 2:
+        score += 0.3  # Multi-mutation = high complexity
+    elif len(complex_muts) == 1:
+        score += 0.15
+
+    # --- Prior treatment mentions (heuristic) ---
+    prior_treatment_keywords = [
+        "prior treatment", "previously treated", "relapsed",
+        "refractory", "second-line", "third-line", "progression",
+        "resistance", "failed", "recurrent",
+    ]
+    text_lower = clinical_text.lower()
+    for kw in prior_treatment_keywords:
+        if kw in text_lower:
+            score += 0.1
+            break
+
+    # Clamp to [0, 1]
+    score = min(score, 1.0)
+
+    # Decision boundary
+    if score >= 0.5:
+        return "complex", score, 2
+    else:
+        return "simple", score, 1
+
+
+# ---------------------------------------------------------------------------
+# Router Node
+# ---------------------------------------------------------------------------
+
+def router_node(state: AgentState) -> Dict[str, Any]:
+    """Classify case complexity and select the appropriate model tier.
+
+    If the user has set ``user_tier_override`` in the state, that
+    takes precedence over the automatic classification.
+
+    Args:
+        state: Current LangGraph state.
+
+    Returns:
+        State update with routing_decision, complexity_score, selected_tier.
+    """
+    clinical_text: str = state.get("clinical_text", "")
+    entities: Dict[str, Any] = state.get("extracted_entities", {})
+    user_override: Optional[int] = state.get("user_tier_override")
+
+    # Run automatic classification
+    decision, score, auto_tier = _classify_complexity(clinical_text, entities)
+
+    # Apply manual override if present
+    if user_override in (1, 2):
+        selected_tier = user_override
+        logger.info(
+            "Manual tier override applied: Tier %d (auto would be Tier %d, score=%.2f)",
+            user_override, auto_tier, score,
+        )
+    else:
+        selected_tier = auto_tier
+        logger.info(
+            "Auto-routing: decision=%s, score=%.2f → Tier %d",
+            decision, score, selected_tier,
+        )
+
+    return {
+        "routing_decision": decision,
+        "complexity_score": round(score, 4),
+        "selected_tier": selected_tier,
+    }

agents/specialist.py

@@ -0,0 +1,206 @@
+"""
+Specialist Node — Tier-adaptive clinical reasoning with Chain-of-Thought.
+
+Design patterns:
+  - Model Tiering: routes to Qwen3.5-9B (fast) or Qwen3.6-27B (deep)
+  - Reflexion: accepts critic feedback for iterative refinement
+  - Anti-Hallucination: system prompt strictly forbids inventing treatments
+
+The specialist produces a structured recommendation with explicit
+reasoning sections (Findings → Staging → Treatment → Recommendation).
+"""
+
+import logging
+from typing import Dict, Any
+
+from .state import AgentState
+from .tools import call_tier_model, get_tier_spec
+
+logger = logging.getLogger(__name__)
+
+
+# ---------------------------------------------------------------------------
+# Prompt Engineering
+# ---------------------------------------------------------------------------
+
+_SYSTEM_PROMPT_TEMPLATE = """\
+You are an expert clinical oncologist operating as part of the OncoAgent system.
+Your task is to analyze the patient case and provide the most appropriate clinical
+next steps based STRICTLY on the provided guidelines.
+
+MODEL TIER: {tier_name} ({tier_description})
+
+DIAGNOSTIC RIGOR POLICY:
+1. You MUST verify if a definitive diagnosis (e.g., pathology report, biopsy) exists.
+2. If diagnostic evidence is missing or inconclusive, your PRIMARY recommendation 
+   MUST be the specific diagnostic procedure needed (e.g., "Esperar informe de biopsia", 
+   "Realizar legrado diagnóstico").
+3. You are STRICTLY FORBIDDEN from assuming cancer exists or jumping to treatment 
+   protocols (surgery, chemo, radiation) if the pathology is not confirmed in the input.
+
+ANTI-HALLUCINATION POLICY:
+1. If the information is NOT explicitly in the guidelines, reply ONLY with: 
+   "Información no concluyente en las guías provistas."
+2. Do NOT invent dosages or protocols.
+
+OUTPUT FORMAT (use this exact structure):
+## Hallazgos Clínicos
+[Summary of current patient presentation]
+
+## Validación Diagnóstica
+[State if pathology/biopsy is present and confirmed. If missing, specify what is needed.]
+
+## Análisis de Estadificación
+[Map findings to staging ONLY if diagnosis is confirmed. Otherwise, state why it's not possible.]
+
+## Opciones de Manejo
+[List clinical next steps or treatment options ONLY if appropriate for the diagnostic stage.]
+
+## Recomendación Final
+[The absolute next step for the clinician with confidence level]
+
+Provide your recommendation in Spanish, clearly citing the guidelines.
+IMPORTANT: Output your recommendation DIRECTLY. Do NOT wrap it in <think> tags."""
+
+
+_USER_PROMPT_TEMPLATE = """\
+Patient Information:
+- Original Text: {clinical_text}
+- Cancer Type: {cancer_type}
+- Stage: {stage}
+- Mutations: {mutations}
+
+Clinical Guidelines Context:
+{context}
+
+{api_evidence}
+
+{critic_feedback_section}
+
+Based ONLY on the guidelines above, what are the recommended clinical next steps?"""
+
+
+def _build_specialist_prompt(
+    state: AgentState,
+) -> tuple[str, str]:
+    """Build the system and user prompts for the specialist.
+
+    Incorporates critic feedback if this is a retry iteration.
+
+    Args:
+        state: Current LangGraph state.
+
+    Returns:
+        Tuple of (system_prompt, user_prompt).
+    """
+    tier = state.get("selected_tier", 1)
+    spec = get_tier_spec(tier)
+
+    system_prompt = _SYSTEM_PROMPT_TEMPLATE.format(
+        tier_name=spec.name,
+        tier_description=spec.description,
+    )
+
+    entities = state.get("extracted_entities", {})
+    context = "\n---\n".join(state.get("rag_context", []))
+    api_evidence = state.get("api_evidence_context", [])
+
+    # Format API evidence if available
+    api_section = ""
+    if api_evidence:
+        api_section = "Additional Evidence (Genomic/Trials):\n" + "\n".join(api_evidence)
+
+    # Inject critic feedback for retry iterations
+    critic_feedback = state.get("critic_feedback", "")
+    critic_attempts = state.get("critic_attempts", 0)
+    feedback_section = ""
+    if critic_attempts > 0 and critic_feedback:
+        feedback_section = (
+            f"\n⚠️ PREVIOUS ATTEMPT FEEDBACK (attempt {critic_attempts}):\n"
+            f"The following issues were identified in your previous recommendation. "
+            f"Please address them in this revision:\n{critic_feedback}\n"
+        )
+
+    user_prompt = _USER_PROMPT_TEMPLATE.format(
+        clinical_text=state.get("clinical_text", ""),
+        cancer_type=entities.get("cancer_type", "Unknown"),
+        stage=entities.get("stage", "Unknown"),
+        mutations=", ".join(entities.get("mutations", [])),
+        context=context,
+        api_evidence=api_section,
+        critic_feedback_section=feedback_section,
+    )
+
+    return system_prompt, user_prompt
+
+
+# ---------------------------------------------------------------------------
+# Specialist Node
+# ---------------------------------------------------------------------------
+
+def specialist_node(state: AgentState) -> Dict[str, Any]:
+    """Generate a clinical recommendation using the tier-adaptive model.
+
+    If critic feedback exists in the state (retry iteration), the feedback
+    is injected into the prompt so the model can self-correct.
+
+    Args:
+        state: Current LangGraph state.
+
+    Returns:
+        State update with clinical_recommendation and reasoning_trace.
+    """
+    context = state.get("rag_context", [])
+    tier = state.get("selected_tier", 1)
+    attempt = state.get("critic_attempts", 0)
+
+    # Guard: no context available
+    if not context:
+        return {
+            "clinical_recommendation": (
+                "Información no concluyente en las guías provistas. "
+                "No se encontró evidencia relevante en la base de datos clínica."
+            ),
+            "reasoning_trace": "No RAG context available — safe fallback triggered.",
+        }
+
+    system_prompt, user_prompt = _build_specialist_prompt(state)
+
+    spec = get_tier_spec(tier)
+    logger.info(
+        "Specialist invoking %s (attempt %d, context chunks: %d)",
+        spec, attempt + 1, len(context),
+    )
+
+    try:
+        recommendation = call_tier_model(
+            tier=tier,
+            system_prompt=system_prompt,
+            user_prompt=user_prompt,
+        )
+
+        # Build reasoning trace for the critic
+        reasoning_trace = (
+            f"Tier: {spec.name} ({spec.model_id})\n"
+            f"Attempt: {attempt + 1}\n"
+            f"Context chunks: {len(context)}\n"
+            f"API evidence items: {len(state.get('api_evidence_context', []))}\n"
+            f"Recommendation length: {len(recommendation)} chars"
+        )
+
+    except RuntimeError as exc:
+        logger.error("Specialist inference failed: %s", exc)
+        recommendation = (
+            "Error en el sistema de inferencia. "
+            "No se pudo generar la recomendación clínica en este momento."
+        )
+        reasoning_trace = f"INFERENCE ERROR: {exc}"
+
+    # Detect if model returned the safe phrase
+    if "información no concluyente" in recommendation.lower():
+        recommendation = "Información no concluyente en las guías provistas."
+
+    return {
+        "clinical_recommendation": recommendation,
+        "reasoning_trace": reasoning_trace,
+    }

agents/state.py

@@ -0,0 +1,105 @@
+"""
+AgentState — Shared state schema for the OncoAgent LangGraph execution.
+
+Design principles (inspired by Claude Code + Hermes Agent):
+  - Immutable input: ``clinical_text`` is never mutated.
+  - Additive outputs: each node writes to its own isolated keys.
+  - Deterministic routing: ``routing_decision`` and ``selected_tier``
+    are set by the Router node using structured logic, never free text.
+  - Per-patient memory: ``patient_id`` isolates session history.
+"""
+
+from typing import TypedDict, Annotated, List, Dict, Any, Optional
+import operator
+from langchain_core.messages import BaseMessage
+from langgraph.graph.message import add_messages
+
+
+class AgentState(TypedDict):
+    """
+    Represents the state of the LangGraph execution for OncoAgent.
+
+    Sections are ordered by the pipeline stage that writes them.
+    Keys prefixed with ``#`` comments indicate which node owns each group.
+    """
+
+    # ------------------------------------------------------------------ #
+    # 0. Session & Patient Context                                       #
+    # ------------------------------------------------------------------ #
+    patient_id: str                  # Unique patient profile ID
+    session_id: str                  # Current session identifier
+    user_tier_override: Optional[int]  # Manual tier override (1 or 2, None = auto)
+    messages: Annotated[List[BaseMessage], add_messages]  # Chat history
+
+    # ------------------------------------------------------------------ #
+    # 1. Input (Immutable — set once at invocation)                      #
+    # ------------------------------------------------------------------ #
+    clinical_text: str
+
+    # ------------------------------------------------------------------ #
+    # 2. Router Node                                                     #
+    # ------------------------------------------------------------------ #
+    routing_decision: str            # "simple" | "complex" | "insufficient"
+    selected_tier: int               # 1 (Qwen 3.5 9B) or 2 (Qwen 3.6 27B)
+    complexity_score: float          # 0.0–1.0 complexity estimate
+
+    # ------------------------------------------------------------------ #
+    # 3. Ingestion Node (PHI clean + entity extraction)                  #
+    # ------------------------------------------------------------------ #
+    extracted_entities: Dict[str, Any]
+    phi_detected: bool
+
+    # ------------------------------------------------------------------ #
+    # 4. Corrective RAG Node                                             #
+    # ------------------------------------------------------------------ #
+    rag_context: List[str]
+    rag_sources: List[str]
+    graph_rag_context: List[str]       # Clinical Knowledge Graph results
+    api_evidence_context: List[str]    # CIViC / ClinicalTrials.gov results
+    rag_confidence: float              # Mean cross-encoder score (0–1)
+    rag_retrieval_count: int           # Results that passed the distance gate
+    rag_grading_pass_count: int        # Documents graded RELEVANT by CRAG
+    rag_query_rewrites: int            # Number of query rewrites performed
+
+    # ------------------------------------------------------------------ #
+    # 5. Specialist Node (Tier-adaptive reasoning)                       #
+    # ------------------------------------------------------------------ #
+    clinical_recommendation: str
+    reasoning_trace: str               # Chain-of-thought breakdown
+
+    # ------------------------------------------------------------------ #
+    # 6. Critic Node (Reflexion loop)                                    #
+    # ------------------------------------------------------------------ #
+    critic_verdict: str                # "PASS" | "FAIL"
+    critic_feedback: str               # Specific issues for specialist retry
+    critic_attempts: int               # Current iteration count (max 2)
+
+    # ------------------------------------------------------------------ #
+    # 7. HITL Gate                                                       #
+    # ------------------------------------------------------------------ #
+    acuity_level: str                  # "low" | "medium" | "high"
+    hitl_required: bool                # True if clinician approval needed
+    hitl_approved: bool                # Set by clinician via UI interrupt
+
+    # ------------------------------------------------------------------ #
+    # 8. Formatter Node (final output)                                   #
+    # ------------------------------------------------------------------ #
+    formatted_recommendation: str      # Markdown-formatted for Gradio
+    confidence_report: Dict[str, Any]  # Full metrics (tier, RAG, critic iters)
+    source_citations: List[str]        # Formatted bibliography
+
+    # ------------------------------------------------------------------ #
+    # 9. Fallback Node                                                   #
+    # ------------------------------------------------------------------ #
+    fallback_reason: str               # Why the system fell back to safe mode
+
+    # ------------------------------------------------------------------ #
+    # 10. Safety (legacy compat + validator output)                      #
+    # ------------------------------------------------------------------ #
+    safety_status: str
+    is_safe: bool
+
+    # ------------------------------------------------------------------ #
+    # 11. Error accumulator (append-only via operator.add)               #
+    # ------------------------------------------------------------------ #
+    errors: Annotated[List[str], operator.add]

agents/tools.py

@@ -0,0 +1,365 @@
+"""
+Shared vLLM client and tier-aware model calling utilities.
+
+All LLM inference across OncoAgent flows through this module,
+ensuring consistent model selection, error handling, and
+environment variable management.
+
+Design inspired by:
+  - Hermes Agent: structured tool calling with JSON output
+  - Claude Code: deterministic harness separating LLM from execution
+
+Production target: AMD Instinct MI300X via ROCm 7.2 + vLLM
+Development fallback: Featherless.ai OpenAI-compatible API
+"""
+
+import os
+import re
+import logging
+from typing import Optional, Dict, Any, List
+from dataclasses import dataclass
+
+from openai import OpenAI
+from dotenv import load_dotenv
+
+load_dotenv()
+
+logger = logging.getLogger(__name__)
+
+
+# ---------------------------------------------------------------------------
+# Tier Configuration
+# ---------------------------------------------------------------------------
+
+@dataclass(frozen=True)
+class TierSpec:
+    """Immutable specification for a model tier."""
+
+    tier_id: int
+    name: str
+    model_id: str
+    description: str
+    max_tokens: int
+    temperature: float
+
+    def __str__(self) -> str:
+        return f"Tier {self.tier_id}: {self.name} ({self.model_id})"
+
+
+# Production tier definitions — Qwen 3.5 / 3.6 as per project rules
+TIER_SPECS: Dict[int, TierSpec] = {
+    1: TierSpec(
+        tier_id=1,
+        name="Speed Triage",
+        model_id=os.getenv("TIER1_MODEL_ID", "Qwen/Qwen3.5-9B"),
+        description="Fast model for initial triage and low-complexity cases.",
+        max_tokens=2048,
+        temperature=0.1,
+    ),
+    2: TierSpec(
+        tier_id=2,
+        name="Deep Reasoning",
+        model_id=os.getenv("TIER2_MODEL_ID", "Qwen/Qwen3.6-27B"),
+        description="High-reasoning model for complex oncology cases and validation.",
+        max_tokens=4096,
+        temperature=0.0,
+    ),
+}
+
+
+# ---------------------------------------------------------------------------
+# Qwen3 Thinking-Mode Handler
+# ---------------------------------------------------------------------------
+
+_THINK_PATTERN = re.compile(r"<think>.*?</think>", re.DOTALL)
+
+
+def _strip_thinking_tokens(text: str) -> str:
+    """Remove Qwen3 <think>...</think> blocks from model output.
+
+    Qwen3 models use an internal reasoning mode that wraps chain-of-thought
+    in <think> tags. We preserve only the final answer for the pipeline.
+
+    Args:
+        text: Raw model output potentially containing <think> blocks.
+
+    Returns:
+        Cleaned text with thinking blocks removed.
+    """
+    cleaned = _THINK_PATTERN.sub("", text).strip()
+    # If everything was inside <think> tags, return the original
+    return cleaned if cleaned else text.strip()
+
+
+# ---------------------------------------------------------------------------
+# vLLM Client Singleton
+# ---------------------------------------------------------------------------
+
+_vllm_client: Optional[OpenAI] = None
+
+
+def get_vllm_client() -> OpenAI:
+    """Return a cached OpenAI-compatible client pointing at vLLM.
+
+    Reads ``VLLM_API_BASE`` and ``VLLM_API_KEY`` from environment.
+
+    Returns:
+        OpenAI client configured for the local vLLM server.
+    """
+    global _vllm_client
+    if _vllm_client is None:
+        api_base = os.getenv("VLLM_API_BASE", "http://localhost:8000/v1")
+        api_key = os.getenv("VLLM_API_KEY", "EMPTY")
+        _vllm_client = OpenAI(base_url=api_base, api_key=api_key)
+        logger.info("vLLM client initialised → %s", api_base)
+    return _vllm_client
+
+
+# ---------------------------------------------------------------------------
+# Model ID Resolution (handles Featherless fallback for dev)
+# ---------------------------------------------------------------------------
+
+def _resolve_model_id(spec: TierSpec) -> str:
+    """Resolve the actual model ID to use for API calls.
+
+    In production (local vLLM), we use the exact model ID.
+    In development (Featherless.ai), some models may not be available,
+    so we check for configured fallbacks.
+
+    Args:
+        spec: The TierSpec for the requested tier.
+
+    Returns:
+        The model ID string to pass to the API.
+    """
+    api_base = os.getenv("VLLM_API_BASE", "http://localhost:8000/v1")
+    is_featherless = "featherless" in api_base.lower()
+
+    if is_featherless and spec.tier_id == 2:
+        # Qwen3.6-27B is not available on Featherless — use fallback
+        fallback = os.getenv("TIER2_FEATHERLESS_FALLBACK", "Qwen/Qwen3.5-27B")
+        logger.info(
+            "Featherless detected: Tier 2 fallback %s → %s",
+            spec.model_id, fallback,
+        )
+        return fallback
+
+    return spec.model_id
+
+
+# ---------------------------------------------------------------------------
+# Local Adapter Manager (PEFT — AMD MI300X only)
+# ---------------------------------------------------------------------------
+
+class LocalModelManager:
+    """Singleton to manage local LoRA model loading and inference.
+
+    Only used on the AMD droplet with working ROCm/GPU drivers.
+    In development without GPU, this is skipped entirely.
+    """
+
+    _instance = None
+
+    def __new__(cls):
+        if cls._instance is None:
+            cls._instance = super(LocalModelManager, cls).__new__(cls)
+            cls._instance.model = None
+            cls._instance.tokenizer = None
+            cls._instance.initialized = False
+        return cls._instance
+
+    def initialize(self) -> None:
+        """Load the base model and LoRA adapters."""
+        if self.initialized:
+            return
+
+        try:
+            from transformers import AutoModelForCausalLM, AutoTokenizer
+            from peft import PeftModel
+            import torch
+        except ImportError:
+            logger.warning("Transformers/PEFT/Torch not installed. Local inference disabled.")
+            return
+
+        adapter_path = os.getenv("LOCAL_ADAPTER_PATH")
+        base_model_id = os.getenv("BASE_MODEL_ID", "Qwen/Qwen3.5-9B")
+
+        if not adapter_path or not os.path.exists(adapter_path):
+            logger.error("Local adapter path not found: %s", adapter_path)
+            return
+
+        logger.info("Loading base model %s + adapters %s...", base_model_id, adapter_path)
+        try:
+            self.tokenizer = AutoTokenizer.from_pretrained(
+                base_model_id, trust_remote_code=True,
+            )
+            base_model = AutoModelForCausalLM.from_pretrained(
+                base_model_id,
+                dtype=torch.bfloat16,
+                device_map="auto",
+                trust_remote_code=True,
+            )
+            self.model = PeftModel.from_pretrained(base_model, adapter_path)
+            self.model.eval()
+            self.initialized = True
+            logger.info("Local BF16 model ready on %s", os.getenv("DEVICE", "cuda"))
+        except Exception as exc:
+            logger.error("Failed to load local model: %s", exc)
+
+    def generate(
+        self,
+        system_prompt: str,
+        user_prompt: str,
+        max_tokens: int,
+        temperature: float,
+    ) -> str:
+        """Run inference using the loaded local model."""
+        if not self.initialized:
+            self.initialize()
+        if not self.initialized:
+            raise RuntimeError("Local model manager not initialized.")
+
+        import torch
+
+        messages = [
+            {"role": "system", "content": system_prompt},
+            {"role": "user", "content": user_prompt},
+        ]
+        prompt_str = self.tokenizer.apply_chat_template(
+            messages, tokenize=False, add_generation_prompt=True,
+        )
+        inputs = self.tokenizer(text=prompt_str, return_tensors="pt").to("cuda")
+
+        with torch.no_grad():
+            outputs = self.model.generate(
+                **inputs,
+                max_new_tokens=max_tokens,
+                temperature=temperature,
+                do_sample=temperature > 0,
+                use_cache=True,
+                pad_token_id=self.tokenizer.pad_token_id,
+            )
+        generated_ids = outputs[:, inputs.input_ids.shape[1]:]
+        response = self.tokenizer.batch_decode(generated_ids, skip_special_tokens=True)[0]
+        return _strip_thinking_tokens(response)
+
+
+_local_manager = LocalModelManager()
+
+
+# ---------------------------------------------------------------------------
+# Tier-Aware Model Calling
+# ---------------------------------------------------------------------------
+
+def call_tier_model(
+    tier: int,
+    system_prompt: str,
+    user_prompt: str,
+    max_tokens: Optional[int] = None,
+    temperature: Optional[float] = None,
+    json_mode: bool = False,
+) -> str:
+    """Call the appropriate model based on the selected tier.
+
+    This is the *single entry point* for all LLM inference in OncoAgent.
+    Every node must call this function instead of instantiating clients.
+
+    Flow:
+      1. If USE_LOCAL_ADAPTERS=true AND tier=1 → try local PEFT inference
+      2. If local fails or not enabled → route through vLLM/Featherless API
+
+    Args:
+        tier: Model tier (1 = fast 9B, 2 = deep 27B).
+        system_prompt: System-level instructions.
+        user_prompt: User-level content / query.
+        max_tokens: Override the tier's default max_tokens.
+        temperature: Override the tier's default temperature.
+        json_mode: If True, request JSON response format.
+
+    Returns:
+        The model's text response (stripped of thinking tokens).
+
+    Raises:
+        ValueError: If the tier is not 1 or 2.
+        RuntimeError: If the vLLM server is unreachable.
+    """
+    spec = TIER_SPECS.get(tier)
+    if spec is None:
+        raise ValueError(f"Invalid tier {tier}. Must be 1 or 2.")
+
+    effective_max_tokens = max_tokens or spec.max_tokens
+    effective_temperature = temperature if temperature is not None else spec.temperature
+
+    logger.info(
+        "Calling %s (max_tokens=%d, temp=%.2f, json=%s)",
+        spec, effective_max_tokens, effective_temperature, json_mode,
+    )
+
+    # --- Path 1: Local LoRA adapters (MI300X only) ---
+    use_local = os.getenv("USE_LOCAL_ADAPTERS", "false").lower() == "true"
+    if tier == 1 and use_local:
+        try:
+            logger.info("Routing Tier 1 to local LoRA adapters...")
+            return _local_manager.generate(
+                system_prompt=system_prompt,
+                user_prompt=user_prompt,
+                max_tokens=effective_max_tokens,
+                temperature=effective_temperature,
+            )
+        except Exception as local_exc:
+            logger.warning("Local inference failed, falling back to API: %s", local_exc)
+
+    # --- Path 2: vLLM / Featherless API ---
+    model_id = _resolve_model_id(spec)
+
+    try:
+        client = get_vllm_client()
+
+        kwargs: Dict[str, Any] = {
+            "model": model_id,
+            "messages": [
+                {"role": "system", "content": system_prompt},
+                {"role": "user", "content": user_prompt},
+            ],
+            "temperature": effective_temperature,
+            "max_tokens": effective_max_tokens,
+        }
+
+        if json_mode:
+            kwargs["response_format"] = {"type": "json_object"}
+
+        response = client.chat.completions.create(**kwargs)
+        raw_text = response.choices[0].message.content or ""
+        text = _strip_thinking_tokens(raw_text)
+
+        if not text:
+            logger.warning(
+                "Model returned empty response (raw_len=%d). "
+                "May be all <think> tokens. Returning raw.",
+                len(raw_text),
+            )
+            text = raw_text.strip() if raw_text else ""
+
+        logger.debug("Response length: %d chars", len(text))
+        return text
+
+    except Exception as exc:
+        logger.error("vLLM call failed for %s: %s", spec, exc)
+        raise RuntimeError(
+            f"Error connecting to vLLM ({model_id}): {exc}"
+        ) from exc
+
+
+def get_tier_spec(tier: int) -> TierSpec:
+    """Retrieve the TierSpec for the given tier number.
+
+    Args:
+        tier: 1 or 2.
+
+    Returns:
+        The corresponding TierSpec.
+    """
+    spec = TIER_SPECS.get(tier)
+    if spec is None:
+        raise ValueError(f"Invalid tier {tier}. Available: {list(TIER_SPECS.keys())}")
+    return spec

app.py

@@ -235,6 +235,74 @@ label, .gr-input-label { color: #94a3b8 !important; }
     padding: 12px; border-top: 1px solid #1e293b;
 }
 
+/* Landing Page */
+.landing-page {
+    display: flex;
+    flex-direction: column;
+    align-items: center;
+    justify-content: center;
+    min-height: 80vh;
+    text-align: center;
+    background: radial-gradient(circle at center, rgba(14, 165, 233, 0.08) 0%, transparent 60%);
+    border-radius: 24px;
+    padding: 40px;
+}
+.hero-title {
+    font-family: 'Figtree', sans-serif;
+    font-size: 3.8rem;
+    font-weight: 800;
+    color: #f8fafc;
+    margin-bottom: 16px;
+    letter-spacing: -0.03em;
+    background: linear-gradient(135deg, #e0f2fe 0%, #38bdf8 100%);
+    -webkit-background-clip: text;
+    -webkit-text-fill-color: transparent;
+}
+.hero-subtitle {
+    font-size: 1.25rem;
+    color: #94a3b8;
+    max-width: 650px;
+    margin: 0 auto 32px auto;
+    line-height: 1.6;
+}
+.btn-launch {
+    background: linear-gradient(135deg, #0ea5e9, #0284c7) !important;
+    border: none !important; color: #fff !important;
+    font-size: 1.15rem !important; font-weight: 600 !important;
+    padding: 16px 42px !important; border-radius: 12px !important;
+    cursor: pointer !important;
+    transition: all 0.2s ease-out !important;
+    box-shadow: 0 8px 24px rgba(14, 165, 233, 0.3) !important;
+}
+.btn-launch:hover {
+    transform: translateY(-2px) !important;
+    box-shadow: 0 12px 32px rgba(14, 165, 233, 0.4) !important;
+}
+.features-grid {
+    display: flex;
+    gap: 24px;
+    margin-top: 56px;
+    justify-content: center;
+    flex-wrap: wrap;
+}
+.feature-card {
+    background: rgba(30, 41, 59, 0.6);
+    border: 1px solid rgba(51, 65, 85, 0.5);
+    border-radius: 16px;
+    padding: 24px;
+    width: 280px;
+    text-align: left;
+    backdrop-filter: blur(12px);
+    transition: transform 0.2s ease, border-color 0.2s ease;
+}
+.feature-card:hover {
+    transform: translateY(-4px);
+    border-color: rgba(14, 165, 233, 0.4);
+}
+.feature-icon { font-size: 2rem; margin-bottom: 14px; }
+.feature-title { color: #f1f5f9; font-weight: 600; margin-bottom: 8px; font-size: 1.1rem; }
+.feature-desc { color: #64748b; font-size: 0.88rem; line-height: 1.5; }
+
 /* Reduced motion */
 @media (prefers-reduced-motion: reduce) {
     *, *::before, *::after {
@@ -562,46 +630,87 @@ with gr.Blocks(
     title="OncoAgent — Oncology Triage Demo",
     theme=gr.themes.Base(),
 ) as demo:
-    # Header
-    gr.HTML(HEADER_HTML)
-    gr.HTML(INFO_HTML)
-
-    # Chat
-    chatbot = gr.Chatbot(
-        type="messages",
-        label="Clinical Triage Chat",
-        height=520,
-        show_label=False,
-        show_copy_button=True,
-        render_markdown=True,
-        elem_classes=["card"],
-    )
-
-    # Controls
-    with gr.Row():
-        with gr.Column(scale=3):
-            txt = gr.Textbox(
-                placeholder="Enter a clinical case or click '▶ View Demo'...",
-                show_label=False,
-                lines=2,
-                max_lines=5,
-            )
-        with gr.Column(scale=1, min_width=180):
-            demo_btn = gr.Button(
-                "▶ View Demo",
-                elem_classes=["btn-demo"],
-                size="lg",
-            )
-
-    with gr.Row():
-        send_btn = gr.Button("Send", elem_classes=["btn-primary"], size="sm")
-        clear_btn = gr.Button("🗑 Clear", variant="secondary", size="sm")
-
-    # Footer
-    gr.HTML(FOOTER_HTML)
+    # ── Landing Page ──────────────────────────────────────────────────
+    with gr.Column(elem_classes=["landing-page"], visible=True) as landing_page:
+        gr.HTML("""
+        <div class="hero-title">🧬 OncoAgent</div>
+        <div class="hero-subtitle">
+            An open-source, multi-agent AI system designed for clinical oncology triage. 
+            Powered by AMD Instinct™ MI300X, LangGraph, and specialized Qwen models.
+        </div>
+        """)
+        
+        launch_btn = gr.Button("🚀 Launch Demo", elem_classes=["btn-launch"], size="lg")
+        
+        gr.HTML("""
+        <div class="features-grid">
+            <div class="feature-card">
+                <div class="feature-icon">📚</div>
+                <div class="feature-title">Corrective RAG</div>
+                <div class="feature-desc">Grounded in 170+ NCCN & ESMO guidelines with distance-gating to prevent hallucinations.</div>
+            </div>
+            <div class="feature-card">
+                <div class="feature-icon">🧠</div>
+                <div class="feature-title">Multi-Agent Reasoning</div>
+                <div class="feature-desc">Tiered architecture (Qwen3.5-9B Router + Qwen3.6-27B Specialist) for complex clinical analysis.</div>
+            </div>
+            <div class="feature-card">
+                <div class="feature-icon">🛡️</div>
+                <div class="feature-title">Clinical Safety</div>
+                <div class="feature-desc">Reflexion loops validate outputs against strict medical criteria before clinician review.</div>
+            </div>
+        </div>
+        """)
+
+    # ── Main App ──────────────────────────────────────────────────────
+    with gr.Column(visible=False) as app_page:
+        # Header
+        gr.HTML(HEADER_HTML)
+        gr.HTML(INFO_HTML)
+
+        # Chat
+        chatbot = gr.Chatbot(
+            type="messages",
+            label="Clinical Triage Chat",
+            height=520,
+            show_label=False,
+            show_copy_button=True,
+            render_markdown=True,
+            elem_classes=["card"],
+        )
+
+        # Controls
+        with gr.Row():
+            with gr.Column(scale=3):
+                txt = gr.Textbox(
+                    placeholder="Enter a clinical case or click '▶ View Demo'...",
+                    show_label=False,
+                    lines=2,
+                    max_lines=5,
+                )
+            with gr.Column(scale=1, min_width=180):
+                demo_btn = gr.Button(
+                    "▶ View Demo",
+                    elem_classes=["btn-demo"],
+                    size="lg",
+                )
+
+        with gr.Row():
+            send_btn = gr.Button("Send", elem_classes=["btn-primary"], size="sm")
+            clear_btn = gr.Button("🗑 Clear", variant="secondary", size="sm")
+
+        # Footer
+        gr.HTML(FOOTER_HTML)
 
     # ── Event Handlers ────────────────────────────────────────────────
 
+    # Landing page navigation
+    launch_btn.click(
+        fn=lambda: [gr.update(visible=False), gr.update(visible=True)],
+        inputs=None,
+        outputs=[landing_page, app_page],
+    )
+
     demo_btn.click(
         fn=run_demo,
         inputs=None,

config.json

@@ -0,0 +1 @@
+{'version': '6.14.0', 'api_prefix': '/gradio_api', 'mode': 'blocks', 'app_id': 4110450775379994823, 'dev_mode': False, 'vibe_mode': False, 'analytics_enabled': True, 'components': [{'id': 1, 'type': 'html', 'props': {'_retryable': False, '_undoable': False, 'likeable': False, 'streamable': False, 'value': '<link rel="stylesheet" href="https://fonts.googleapis.com/css2?family=Figtree:wght@400;500;600;700&family=Inter:wght@300;400;500;600&display=swap">', 'html_template': '${value}', 'css_template': '', 'js_on_load': "element.addEventListener('click', function() { trigger('click') });", 'apply_default_css': True, 'show_label': False, 'visible': True, 'elem_classes': [], 'preserved_by_key': ['value'], 'container': False, 'padding': False, 'autoscroll': False, 'buttons': [], 'props': {}, 'name': 'html', '_selectable': False, 'component_class_name': 'HTML'}, 'skip_api': False, 'component_class_id': 'af5f63fae9620e1007439226451e1707e9a6016bcb073dc5a0b6433126cda1fc', 'key': None, 'api_info': {'type': 'string'}, 'api_info_as_input': {'type': 'string'}, 'api_info_as_output': {'type': 'string'}, 'example_inputs': '<p>Hello</p>'}, {'id': 2, 'type': 'html', 'props': {'_retryable': False, '_undoable': False, 'likeable': False, 'streamable': False, 'value': "<div class='header-bar'><span class='brand-name'>OncoAgent</span><span class='hw-badge'>AMD Instinct MI300X</span></div>", 'html_template': '${value}', 'css_template': '', 'js_on_load': "element.addEventListener('click', function() { trigger('click') });", 'apply_default_css': True, 'show_label': False, 'visible': True, 'elem_classes': [], 'preserved_by_key': ['value'], 'container': False, 'padding': False, 'autoscroll': False, 'buttons': [], 'props': {}, 'name': 'html', '_selectable': False, 'component_class_name': 'HTML'}, 'skip_api': False, 'component_class_id': 'af5f63fae9620e1007439226451e1707e9a6016bcb073dc5a0b6433126cda1fc', 'key': None, 'api_info': {'type': 'string'}, 'api_info_as_input': {'type': 'string'}, 'api_info_as_output': {'type': 'string'}, 'example_inputs': '<p>Hello</p>'}, {'id': 3, 'type': 'row', 'props': {'variant': 'default', 'visible': True, 'elem_classes': [], 'equal_height': False, 'show_progress': False, 'preserved_by_key': [], 'name': 'row'}, 'skip_api': True, 'component_class_id': '6f8a6130c432a547a95664f7f2f2a01fd8019e8e9b05282a61688092ea105a01', 'key': None}, {'id': 4, 'type': 'column', 'props': {'scale': 1, 'min_width': 280, 'variant': 'default', 'visible': True, 'elem_classes': [], 'show_progress': False, 'preserved_by_key': [], 'name': 'column'}, 'skip_api': True, 'component_class_id': '0fb38a7a679d0444705b8d6520b557d7fcfac8cb0bd868952cdb1503899457a3', 'key': None}, {'id': 5, 'type': 'column', 'props': {'scale': 1, 'min_width': 320, 'variant': 'default', 'visible': True, 'elem_classes': ['card'], 'show_progress': False, 'preserved_by_key': [], 'name': 'column'}, 'skip_api': True, 'component_class_id': '0fb38a7a679d0444705b8d6520b557d7fcfac8cb0bd868952cdb1503899457a3', 'key': None}, {'id': 6, 'type': 'html', 'props': {'_retryable': False, '_undoable': False, 'likeable': False, 'streamable': False, 'value': "<div class='section-title'>Session</div>", 'html_template': '${value}', 'css_template': '', 'js_on_load': "element.addEventListener('click', function() { trigger('click') });", 'apply_default_css': True, 'show_label': False, 'visible': True, 'elem_classes': [], 'preserved_by_key': ['value'], 'container': False, 'padding': False, 'autoscroll': False, 'buttons': [], 'props': {}, 'name': 'html', '_selectable': False, 'component_class_name': 'HTML'}, 'skip_api': False, 'component_class_id': 'af5f63fae9620e1007439226451e1707e9a6016bcb073dc5a0b6433126cda1fc', 'key': None, 'api_info': {'type': 'string'}, 'api_info_as_input': {'type': 'string'}, 'api_info_as_output': {'type': 'string'}, 'example_inputs': '<p>Hello</p>'}, {'id': 7, 'type': 'textbox', 'props': {'value': 'PT-5659', 'type': 'text', 'lines': 1, 'label': 'Patient ID', 'info': 'Unique session for memory isolation', 'show_label': True, 'container': True, 'min_width': 160, 'interactive': True, 'visible': True, 'autofocus': False, 'autoscroll': True, 'elem_classes': [], 'preserved_by_key': ['value'], 'rtl': False, 'buttons': [], 'submit_btn': False, 'stop_btn': False, 'name': 'textbox', '_selectable': False}, 'skip_api': False, 'component_class_id': 'de54b6fd6ce8622ae36d11c1cbd43b965ca46f0c1fe283a7cec562dcb91a3208', 'key': None, 'api_info': {'type': 'string'}, 'api_info_as_input': {'type': 'string'}, 'api_info_as_output': {'type': 'string'}, 'example_inputs': 'Hello!!'}, {'id': 8, 'type': 'dropdown', 'props': {'choices': [('auto', 'auto'), ('9b', '9b'), ('27b', '27b')], 'value': 'auto', 'type': 'value', 'allow_custom_value': False, 'filterable': True, 'label': 'Model Tier', 'info': 'Auto-routes based on case complexity', 'show_label': True, 'container': True, 'min_width': 160, 'visible': True, 'elem_classes': [], 'preserved_by_key': ['value'], 'buttons': [], 'name': 'dropdown', '_selectable': False}, 'skip_api': False, 'component_class_id': '2e97f31499001547c4b27c45798eb5990573be2b0b8f11854c47c3b044b07aab', 'key': None, 'api_info': {'type': 'string', 'enum': ['auto', '9b', '27b']}, 'api_info_as_input': {'type': 'string', 'enum': ['auto', '9b', '27b']}, 'api_info_as_output': {'type': 'string', 'enum': ['auto', '9b', '27b']}, 'example_inputs': 'auto'}, {'id': 9, 'type': 'form', 'props': {'scale': 0, 'min_width': 0, 'preserved_by_key': [], 'name': 'form'}, 'skip_api': True, 'component_class_id': 'aa68c082b9d7ec5a5fe2959e3c704e535742029261408360fcdbad9c6db4eb31', 'key': None}, {'id': 10, 'type': 'row', 'props': {'variant': 'default', 'visible': True, 'elem_classes': [], 'equal_height': False, 'show_progress': False, 'preserved_by_key': [], 'name': 'row'}, 'skip_api': True, 'component_class_id': '6f8a6130c432a547a95664f7f2f2a01fd8019e8e9b05282a61688092ea105a01', 'key': None}, {'id': 11, 'type': 'column', 'props': {'scale': 1, 'min_width': 100, 'variant': 'default', 'visible': True, 'elem_classes': ['kpi-tile'], 'show_progress': False, 'preserved_by_key': [], 'name': 'column'}, 'skip_api': True, 'component_class_id': '0fb38a7a679d0444705b8d6520b557d7fcfac8cb0bd868952cdb1503899457a3', 'key': None}, {'id': 12, 'type': 'html', 'props': {'_retryable': False, '_undoable': False, 'likeable': False, 'streamable': False, 'value': "<div class='kpi-label'>Confidence</div><div class='kpi-value' id='kpi-confidence'>—</div>", 'html_template': '${value}', 'css_template': '', 'js_on_load': "element.addEventListener('click', function() { trigger('click') });", 'apply_default_css': True, 'show_label': False, 'visible': True, 'elem_classes': [], 'preserved_by_key': ['value'], 'container': False, 'padding': False, 'autoscroll': False, 'buttons': [], 'props': {}, 'name': 'html', '_selectable': False, 'component_class_name': 'HTML'}, 'skip_api': False, 'component_class_id': 'af5f63fae9620e1007439226451e1707e9a6016bcb073dc5a0b6433126cda1fc', 'key': None, 'api_info': {'type': 'string'}, 'api_info_as_input': {'type': 'string'}, 'api_info_as_output': {'type': 'string'}, 'example_inputs': '<p>Hello</p>'}, {'id': 13, 'type': 'label', 'props': {'value': {}, 'label': 'Confidence', 'show_label': True, 'container': True, 'min_width': 160, 'visible': False, 'elem_classes': [], 'preserved_by_key': ['value'], 'show_heading': True, 'buttons': [], 'name': 'label', '_selectable': False}, 'skip_api': False, 'component_class_id': 'e06edc5731c4fc699cba51f556272631f237cf9443d7c1bb4d06c4b7556aa63a', 'key': None, 'api_info': {'$defs': {'LabelConfidence': {'properties': {'label': {'anyOf': [{'type': 'string'}, {'type': 'integer'}, {'type': 'number'}, {'type': 'null'}], 'default': None, 'title': 'Label'}, 'confidence': {'anyOf': [{'type': 'number'}, {'type': 'null'}], 'default': None, 'title': 'Confidence'}}, 'title': 'LabelConfidence', 'type': 'object'}}, 'properties': {'label': {'anyOf': [{'type': 'string'}, {'type': 'integer'}, {'type': 'number'}, {'type': 'null'}], 'default': None, 'title': 'Label'}, 'confidences': {'anyOf': [{'items': {'$ref': '#/$defs/LabelConfidence'}, 'type': 'array'}, {'type': 'null'}], 'default': None, 'title': 'Confidences'}}, 'title': 'LabelData', 'type': 'object', 'additional_description': None}, 'api_info_as_input': {'$defs': {'LabelConfidence': {'properties': {'label': {'anyOf': [{'type': 'string'}, {'type': 'integer'}, {'type': 'number'}, {'type': 'null'}], 'default': None, 'title': 'Label'}, 'confidence': {'anyOf': [{'type': 'number'}, {'type': 'null'}], 'default': None, 'title': 'Confidence'}}, 'title': 'LabelConfidence', 'type': 'object'}}, 'properties': {'label': {'anyOf': [{'type': 'string'}, {'type': 'integer'}, {'type': 'number'}, {'type': 'null'}], 'default': None, 'title': 'Label'}, 'confidences': {'anyOf': [{'items': {'$ref': '#/$defs/LabelConfidence'}, 'type': 'array'}, {'type': 'null'}], 'default': None, 'title': 'Confidences'}}, 'title': 'LabelData', 'type': 'object', 'additional_description': None}, 'api_info_as_output': {'$defs': {'LabelConfidence': {'properties': {'label': {'anyOf': [{'type': 'string'}, {'type': 'integer'}, {'type': 'number'}, {'type': 'null'}], 'default': None, 'title': 'Label'}, 'confidence': {'anyOf': [{'type': 'number'}, {'type': 'null'}], 'default': None, 'title': 'Confidence'}}, 'title': 'LabelConfidence', 'type': 'object'}}, 'properties': {'label': {'anyOf': [{'type': 'string'}, {'type': 'integer'}, {'type': 'number'}, {'type': 'null'}], 'default': None, 'title': 'Label'}, 'confidences': {'anyOf': [{'items': {'$ref': '#/$defs/LabelConfidence'}, 'type': 'array'}, {'type': 'null'}], 'default': None, 'title': 'Confidences'}}, 'title': 'LabelData', 'type': 'object', 'additional_description': None}, 'example_inputs': {'label': 'Cat', 'confidences': [{'label': 'cat', 'confidence': 0.9}, {'label': 'dog', 'confidence': 0.1}]}}, {'id': 14, 'type': 'column', 'props': {'scale': 1, 'min_width': 100, 'variant': 'default', 'visible': True, 'elem_classes': ['kpi-tile'], 'show_progress': False, 'preserved_by_key': [], 'name': 'column'}, 'skip_api': True, 'component_class_id': '0fb38a7a679d0444705b8d6520b557d7fcfac8cb0bd868952cdb1503899457a3', 'key': None}, {'id': 15, 'type': 'html', 'props': {'_retryable': False, '_undoable': False, 'likeable': False, 'streamable': False, 'value': "<div class='kpi-label'>Sources</div><div class='kpi-value' id='kpi-sources'>—</div>", 'html_template': '${value}', 'css_template': '', 'js_on_load': "element.addEventListener('click', function() { trigger('click') });", 'apply_default_css': True, 'show_label': False, 'visible': True, 'elem_classes': [], 'preserved_by_key': ['value'], 'container': False, 'padding': False, 'autoscroll': False, 'buttons': [], 'props': {}, 'name': 'html', '_selectable': False, 'component_class_name': 'HTML'}, 'skip_api': False, 'component_class_id': 'af5f63fae9620e1007439226451e1707e9a6016bcb073dc5a0b6433126cda1fc', 'key': None, 'api_info': {'type': 'string'}, 'api_info_as_input': {'type': 'string'}, 'api_info_as_output': {'type': 'string'}, 'example_inputs': '<p>Hello</p>'}, {'id': 16, 'type': 'label', 'props': {'value': {}, 'label': 'Sources', 'show_label': True, 'container': True, 'min_width': 160, 'visible': False, 'elem_classes': [], 'preserved_by_key': ['value'], 'show_heading': True, 'buttons': [], 'name': 'label', '_selectable': False}, 'skip_api': False, 'component_class_id': 'e06edc5731c4fc699cba51f556272631f237cf9443d7c1bb4d06c4b7556aa63a', 'key': None, 'api_info': {'$defs': {'LabelConfidence': {'properties': {'label': {'anyOf': [{'type': 'string'}, {'type': 'integer'}, {'type': 'number'}, {'type': 'null'}], 'default': None, 'title': 'Label'}, 'confidence': {'anyOf': [{'type': 'number'}, {'type': 'null'}], 'default': None, 'title': 'Confidence'}}, 'title': 'LabelConfidence', 'type': 'object'}}, 'properties': {'label': {'anyOf': [{'type': 'string'}, {'type': 'integer'}, {'type': 'number'}, {'type': 'null'}], 'default': None, 'title': 'Label'}, 'confidences': {'anyOf': [{'items': {'$ref': '#/$defs/LabelConfidence'}, 'type': 'array'}, {'type': 'null'}], 'default': None, 'title': 'Confidences'}}, 'title': 'LabelData', 'type': 'object', 'additional_description': None}, 'api_info_as_input': {'$defs': {'LabelConfidence': {'properties': {'label': {'anyOf': [{'type': 'string'}, {'type': 'integer'}, {'type': 'number'}, {'type': 'null'}], 'default': None, 'title': 'Label'}, 'confidence': {'anyOf': [{'type': 'number'}, {'type': 'null'}], 'default': None, 'title': 'Confidence'}}, 'title': 'LabelConfidence', 'type': 'object'}}, 'properties': {'label': {'anyOf': [{'type': 'string'}, {'type': 'integer'}, {'type': 'number'}, {'type': 'null'}], 'default': None, 'title': 'Label'}, 'confidences': {'anyOf': [{'items': {'$ref': '#/$defs/LabelConfidence'}, 'type': 'array'}, {'type': 'null'}], 'default': None, 'title': 'Confidences'}}, 'title': 'LabelData', 'type': 'object', 'additional_description': None}, 'api_info_as_output': {'$defs': {'LabelConfidence': {'properties': {'label': {'anyOf': [{'type': 'string'}, {'type': 'integer'}, {'type': 'number'}, {'type': 'null'}], 'default': None, 'title': 'Label'}, 'confidence': {'anyOf': [{'type': 'number'}, {'type': 'null'}], 'default': None, 'title': 'Confidence'}}, 'title': 'LabelConfidence', 'type': 'object'}}, 'properties': {'label': {'anyOf': [{'type': 'string'}, {'type': 'integer'}, {'type': 'number'}, {'type': 'null'}], 'default': None, 'title': 'Label'}, 'confidences': {'anyOf': [{'items': {'$ref': '#/$defs/LabelConfidence'}, 'type': 'array'}, {'type': 'null'}], 'default': None, 'title': 'Confidences'}}, 'title': 'LabelData', 'type': 'object', 'additional_description': None}, 'example_inputs': {'label': 'Cat', 'confidences': [{'label': 'cat', 'confidence': 0.9}, {'label': 'dog', 'confidence': 0.1}]}}, {'id': 17, 'type': 'tabs', 'props': {'visible': True, 'elem_classes': ['card'], 'preserved_by_key': [], 'name': 'tabs'}, 'skip_api': True, 'component_class_id': 'e3f75337ef6662053675c6300cd49c9ac252225778c27231c96a030da41e3a85', 'key': None}, {'id': 18, 'type': 'tabitem', 'props': {'label': 'Guidelines', 'visible': True, 'interactive': True, 'elem_classes': [], 'preserved_by_key': [], 'render_children': False, 'name': 'tab'}, 'skip_api': True, 'component_class_id': '0989c1da084cf64f1899851572faada48fa5efed022fd0d58d24f61eb6d9142d', 'key': None}, {'id': 19, 'type': 'markdown', 'props': {'value': 'NCCN and ESMO guideline evidence will appear here.', 'show_label': False, 'rtl': False, 'latex_delimiters': [{'left': '$$', 'right': '$$', 'display': True}], 'visible': True, 'elem_classes': [], 'preserved_by_key': ['value'], 'sanitize_html': True, 'line_breaks': False, 'header_links': False, 'container': False, 'padding': False, 'name': 'markdown', '_selectable': False}, 'skip_api': False, 'component_class_id': 'fd30cc896cca4ca584ffd40aa2efc60d9820bc905902b7c1d2b0c228f8aa7654', 'key': None, 'api_info': {'type': 'string'}, 'api_info_as_input': {'type': 'string'}, 'api_info_as_output': {'type': 'string'}, 'example_inputs': '# Hello!'}, {'id': 20, 'type': 'tabitem', 'props': {'label': 'Knowledge Graph', 'visible': True, 'interactive': True, 'elem_classes': [], 'preserved_by_key': [], 'render_children': False, 'name': 'tab'}, 'skip_api': True, 'component_class_id': '0989c1da084cf64f1899851572faada48fa5efed022fd0d58d24f61eb6d9142d', 'key': None}, {'id': 21, 'type': 'markdown', 'props': {'value': 'Knowledge graph connections will appear here.', 'show_label': False, 'rtl': False, 'latex_delimiters': [{'left': '$$', 'right': '$$', 'display': True}], 'visible': True, 'elem_classes': [], 'preserved_by_key': ['value'], 'sanitize_html': True, 'line_breaks': False, 'header_links': False, 'container': False, 'padding': False, 'name': 'markdown', '_selectable': False}, 'skip_api': False, 'component_class_id': 'fd30cc896cca4ca584ffd40aa2efc60d9820bc905902b7c1d2b0c228f8aa7654', 'key': None, 'api_info': {'type': 'string'}, 'api_info_as_input': {'type': 'string'}, 'api_info_as_output': {'type': 'string'}, 'example_inputs': '# Hello!'}, {'id': 22, 'type': 'tabitem', 'props': {'label': 'API Evidence', 'visible': True, 'interactive': True, 'elem_classes': [], 'preserved_by_key': [], 'render_children': False, 'name': 'tab'}, 'skip_api': True, 'component_class_id': '0989c1da084cf64f1899851572faada48fa5efed022fd0d58d24f61eb6d9142d', 'key': None}, {'id': 23, 'type': 'markdown', 'props': {'value': 'Real-time data from CIViC and ClinicalTrials.gov.', 'show_label': False, 'rtl': False, 'latex_delimiters': [{'left': '$$', 'right': '$$', 'display': True}], 'visible': True, 'elem_classes': [], 'preserved_by_key': ['value'], 'sanitize_html': True, 'line_breaks': False, 'header_links': False, 'container': False, 'padding': False, 'name': 'markdown', '_selectable': False}, 'skip_api': False, 'component_class_id': 'fd30cc896cca4ca584ffd40aa2efc60d9820bc905902b7c1d2b0c228f8aa7654', 'key': None, 'api_info': {'type': 'string'}, 'api_info_as_input': {'type': 'string'}, 'api_info_as_output': {'type': 'string'}, 'example_inputs': '# Hello!'}, {'id': 24, 'type': 'column', 'props': {'scale': 1, 'min_width': 320, 'variant': 'default', 'visible': True, 'elem_classes': ['card'], 'show_progress': False, 'preserved_by_key': [], 'name': 'column'}, 'skip_api': True, 'component_class_id': '0fb38a7a679d0444705b8d6520b557d7fcfac8cb0bd868952cdb1503899457a3', 'key': None}, {'id': 25, 'type': 'html', 'props': {'_retryable': False, '_undoable': False, 'likeable': False, 'streamable': False, 'value': "<div class='section-title'>System Status</div>", 'html_template': '${value}', 'css_template': '', 'js_on_load': "element.addEventListener('click', function() { trigger('click') });", 'apply_default_css': True, 'show_label': False, 'visible': True, 'elem_classes': [], 'preserved_by_key': ['value'], 'container': False, 'padding': False, 'autoscroll': False, 'buttons': [], 'props': {}, 'name': 'html', '_selectable': False, 'component_class_name': 'HTML'}, 'skip_api': False, 'component_class_id': 'af5f63fae9620e1007439226451e1707e9a6016bcb073dc5a0b6433126cda1fc', 'key': None, 'api_info': {'type': 'string'}, 'api_info_as_input': {'type': 'string'}, 'api_info_as_output': {'type': 'string'}, 'example_inputs': '<p>Hello</p>'}, {'id': 26, 'type': 'markdown', 'props': {'value': "<div class='status-bar'>System ready.</div>", 'show_label': False, 'rtl': False, 'latex_delimiters': [{'left': '$$', 'right': '$$', 'display': True}], 'visible': True, 'elem_id': 'status-box', 'elem_classes': [], 'preserved_by_key': ['value'], 'sanitize_html': True, 'line_breaks': False, 'header_links': False, 'container': False, 'padding': False, 'name': 'markdown', '_selectable': False}, 'skip_api': False, 'component_class_id': 'fd30cc896cca4ca584ffd40aa2efc60d9820bc905902b7c1d2b0c228f8aa7654', 'key': None, 'api_info': {'type': 'string'}, 'api_info_as_input': {'type': 'string'}, 'api_info_as_output': {'type': 'string'}, 'example_inputs': '# Hello!'}, {'id': 27, 'type': 'column', 'props': {'scale': 3, 'min_width': 320, 'variant': 'default', 'visible': True, 'elem_classes': [], 'show_progress': False, 'preserved_by_key': [], 'name': 'column'}, 'skip_api': True, 'component_class_id': '0fb38a7a679d0444705b8d6520b557d7fcfac8cb0bd868952cdb1503899457a3', 'key': None}, {'id': 28, 'type': 'column', 'props': {'scale': 1, 'min_width': 600, 'variant': 'default', 'visible': True, 'elem_classes': ['card'], 'show_progress': False, 'preserved_by_key': [], 'name': 'column'}, 'skip_api': True, 'component_class_id': '0fb38a7a679d0444705b8d6520b557d7fcfac8cb0bd868952cdb1503899457a3', 'key': None}, {'id': 29, 'type': 'chatbot', 'props': {'_undoable': False, '_retryable': False, 'likeable': False, 'value': [], 'label': 'OncoAgent', 'show_label': False, 'container': True, 'min_width': 160, 'visible': True, 'elem_classes': ['gr-chatbot'], 'autoscroll': True, 'preserved_by_key': ['value'], 'height': 620, 'resizable': False, 'latex_delimiters': [{'left': '$$', 'right': '$$', 'display': True}], 'rtl': False, 'buttons': ['share', 'copy', 'copy_all'], 'avatar_images': [None, None], 'sanitize_html': True, 'render_markdown': True, 'feedback_options': ['Like', 'Dislike'], 'line_breaks': True, 'allow_file_downloads': True, 'group_consecutive_messages': True, 'allow_tags': True, 'like_user_message': False, 'name': 'chatbot', '_selectable': False}, 'skip_api': False, 'component_class_id': '6f1aa2d606e01e101fc71adbda61e826828d4b1a485f1b6c746327585193dfb9', 'key': None, 'api_info': {'$defs': {'ComponentMessage': {'properties': {'component': {'title': 'Component', 'type': 'string'}, 'value': {'title': 'Value'}, 'constructor_args': {'additionalProperties': True, 'title': 'Constructor Args', 'type': 'object'}, 'props': {'additionalProperties': True, 'title': 'Props', 'type': 'object'}, 'type': {'const': 'component', 'default': 'component', 'title': 'Type', 'type': 'string'}}, 'required': ['component', 'value', 'constructor_args', 'props'], 'title': 'ComponentMessage', 'type': 'object'}, 'FileData': {'description': 'The FileData class is a subclass of the GradioModel class that represents a file object within a Gradio interface. It is used to store file data and metadata when a file is uploaded.\n\nAttributes:\n    path: The server file path where the file is stored.\n    url: The normalized server URL pointing to the file.\n    size: The size of the file in bytes.\n    orig_name: The original filename before upload.\n    mime_type: The MIME type of the file.\n    is_stream: Indicates whether the file is a stream.\n    meta: Additional metadata used internally (should not be changed).', 'properties': {'path': {'title': 'Path', 'type': 'string'}, 'url': {'anyOf': [{'type': 'string'}, {'type': 'null'}], 'default': None, 'title': 'Url'}, 'size': {'anyOf': [{'type': 'integer'}, {'type': 'null'}], 'default': None, 'title': 'Size'}, 'orig_name': {'anyOf': [{'type': 'string'}, {'type': 'null'}], 'default': None, 'title': 'Orig Name'}, 'mime_type': {'anyOf': [{'type': 'string'}, {'type': 'null'}], 'default': None, 'title': 'Mime Type'}, 'is_stream': {'default': False, 'title': 'Is Stream', 'type': 'boolean'}, 'meta': {'$ref': '#/$defs/FileDataMeta'}}, 'required': ['path'], 'title': 'FileData', 'type': 'object'}, 'FileDataMeta': {'properties': {'_type': {'const': 'gradio.FileData', 'title': 'Type', 'type': 'string'}}, 'required': ['_type'], 'title': 'FileDataMeta', 'type': 'object'}, 'FileMessage': {'properties': {'file': {'$ref': '#/$defs/FileData'}, 'alt_text': {'anyOf': [{'type': 'string'}, {'type': 'null'}], 'default': None, 'title': 'Alt Text'}, 'type': {'const': 'file', 'default': 'file', 'title': 'Type', 'type': 'string'}}, 'required': ['file'], 'title': 'FileMessage', 'type': 'object'}, 'Message': {'properties': {'role': {'title': 'Role', 'type': 'string'}, 'metadata': {'anyOf': [{'$ref': '#/$defs/MetadataDict'}, {'type': 'null'}], 'default': None}, 'content': {'items': {'anyOf': [{'$ref': '#/$defs/TextMessage'}, {'$ref': '#/$defs/FileMessage'}, {'$ref': '#/$defs/ComponentMessage'}]}, 'title': 'Content', 'type': 'array'}, 'options': {'anyOf': [{'items': {'$ref': '#/$defs/OptionDict'}, 'type': 'array'}, {'type': 'null'}], 'default': None, 'title': 'Options'}}, 'required': ['role', 'content'], 'title': 'Message', 'type': 'object'}, 'MetadataDict': {'description': 'A typed dictionary to represent metadata for a message in the Chatbot component. An\ninstance of this dictionary is used for the `metadata` field in a ChatMessage when\nthe chat message should be displayed as a thought.\nParameters:\n    title: The title of the "thought" message. Required if the message is to be displayed as a thought.\n    id: The ID of the message. Only used for nested thoughts. Nested thoughts can be nested by setting the parent_id to the id of the parent thought.\n    parent_id: The ID of the parent message. Only used for nested thoughts.\n    log: A string message to display next to the thought title in a subdued font.\n    duration: The duration of the message in seconds. Appears next to the thought title in a subdued font inside a parentheses.\n    status: if set to `"pending"`, a spinner appears next to the thought title and the accordion is initialized open.  If `status` is `"done"`, the thought accordion is initialized closed. If `status` is not provided, the thought accordion is initialized open and no spinner is displayed.', 'properties': {'title': {'title': 'Title', 'type': 'string'}, 'id': {'anyOf': [{'type': 'integer'}, {'type': 'string'}], 'title': 'Id'}, 'parent_id': {'anyOf': [{'type': 'integer'}, {'type': 'string'}], 'title': 'Parent Id'}, 'log': {'title': 'Log', 'type': 'string'}, 'duration': {'title': 'Duration', 'type': 'number'}, 'status': {'enum': ['pending', 'done'], 'title': 'Status', 'type': 'string'}}, 'title': 'MetadataDict', 'type': 'object'}, 'OptionDict': {'description': 'A typed dictionary to represent an option in a ChatMessage. A list of these\ndictionaries is used for the `options` field in a ChatMessage.\nParameters:\n    value: The value to return when the option is selected.\n    label: The text to display in the option, if different from the value.', 'properties': {'value': {'title': 'Value', 'type': 'string'}, 'label': {'title': 'Label', 'type': 'string'}}, 'required': ['value'], 'title': 'OptionDict', 'type': 'object'}, 'TextMessage': {'properties': {'text': {'title': 'Text', 'type': 'string'}, 'type': {'const': 'text', 'default': 'text', 'title': 'Type', 'type': 'string'}}, 'required': ['text'], 'title': 'TextMessage', 'type': 'object'}}, 'items': {'$ref': '#/$defs/Message'}, 'title': 'ChatbotDataMessages', 'type': 'array', 'additional_description': None}, 'api_info_as_input': {'$defs': {'ComponentMessage': {'properties': {'component': {'title': 'Component', 'type': 'string'}, 'value': {'title': 'Value'}, 'constructor_args': {'additionalProperties': True, 'title': 'Constructor Args', 'type': 'object'}, 'props': {'additionalProperties': True, 'title': 'Props', 'type': 'object'}, 'type': {'const': 'component', 'default': 'component', 'title': 'Type', 'type': 'string'}}, 'required': ['component', 'value', 'constructor_args', 'props'], 'title': 'ComponentMessage', 'type': 'object'}, 'FileData': {'description': 'The FileData class is a subclass of the GradioModel class that represents a file object within a Gradio interface. It is used to store file data and metadata when a file is uploaded.\n\nAttributes:\n    path: The server file path where the file is stored.\n    url: The normalized server URL pointing to the file.\n    size: The size of the file in bytes.\n    orig_name: The original filename before upload.\n    mime_type: The MIME type of the file.\n    is_stream: Indicates whether the file is a stream.\n    meta: Additional metadata used internally (should not be changed).', 'properties': {'path': {'title': 'Path', 'type': 'string'}, 'url': {'anyOf': [{'type': 'string'}, {'type': 'null'}], 'default': None, 'title': 'Url'}, 'size': {'anyOf': [{'type': 'integer'}, {'type': 'null'}], 'default': None, 'title': 'Size'}, 'orig_name': {'anyOf': [{'type': 'string'}, {'type': 'null'}], 'default': None, 'title': 'Orig Name'}, 'mime_type': {'anyOf': [{'type': 'string'}, {'type': 'null'}], 'default': None, 'title': 'Mime Type'}, 'is_stream': {'default': False, 'title': 'Is Stream', 'type': 'boolean'}, 'meta': {'$ref': '#/$defs/FileDataMeta'}}, 'required': ['path'], 'title': 'FileData', 'type': 'object'}, 'FileDataMeta': {'properties': {'_type': {'const': 'gradio.FileData', 'title': 'Type', 'type': 'string'}}, 'required': ['_type'], 'title': 'FileDataMeta', 'type': 'object'}, 'FileMessage': {'properties': {'file': {'$ref': '#/$defs/FileData'}, 'alt_text': {'anyOf': [{'type': 'string'}, {'type': 'null'}], 'default': None, 'title': 'Alt Text'}, 'type': {'const': 'file', 'default': 'file', 'title': 'Type', 'type': 'string'}}, 'required': ['file'], 'title': 'FileMessage', 'type': 'object'}, 'Message': {'properties': {'role': {'title': 'Role', 'type': 'string'}, 'metadata': {'anyOf': [{'$ref': '#/$defs/MetadataDict'}, {'type': 'null'}], 'default': None}, 'content': {'items': {'anyOf': [{'$ref': '#/$defs/TextMessage'}, {'$ref': '#/$defs/FileMessage'}, {'$ref': '#/$defs/ComponentMessage'}]}, 'title': 'Content', 'type': 'array'}, 'options': {'anyOf': [{'items': {'$ref': '#/$defs/OptionDict'}, 'type': 'array'}, {'type': 'null'}], 'default': None, 'title': 'Options'}}, 'required': ['role', 'content'], 'title': 'Message', 'type': 'object'}, 'MetadataDict': {'description': 'A typed dictionary to represent metadata for a message in the Chatbot component. An\ninstance of this dictionary is used for the `metadata` field in a ChatMessage when\nthe chat message should be displayed as a thought.\nParameters:\n    title: The title of the "thought" message. Required if the message is to be displayed as a thought.\n    id: The ID of the message. Only used for nested thoughts. Nested thoughts can be nested by setting the parent_id to the id of the parent thought.\n    parent_id: The ID of the parent message. Only used for nested thoughts.\n    log: A string message to display next to the thought title in a subdued font.\n    duration: The duration of the message in seconds. Appears next to the thought title in a subdued font inside a parentheses.\n    status: if set to `"pending"`, a spinner appears next to the thought title and the accordion is initialized open.  If `status` is `"done"`, the thought accordion is initialized closed. If `status` is not provided, the thought accordion is initialized open and no spinner is displayed.', 'properties': {'title': {'title': 'Title', 'type': 'string'}, 'id': {'anyOf': [{'type': 'integer'}, {'type': 'string'}], 'title': 'Id'}, 'parent_id': {'anyOf': [{'type': 'integer'}, {'type': 'string'}], 'title': 'Parent Id'}, 'log': {'title': 'Log', 'type': 'string'}, 'duration': {'title': 'Duration', 'type': 'number'}, 'status': {'enum': ['pending', 'done'], 'title': 'Status', 'type': 'string'}}, 'title': 'MetadataDict', 'type': 'object'}, 'OptionDict': {'description': 'A typed dictionary to represent an option in a ChatMessage. A list of these\ndictionaries is used for the `options` field in a ChatMessage.\nParameters:\n    value: The value to return when the option is selected.\n    label: The text to display in the option, if different from the value.', 'properties': {'value': {'title': 'Value', 'type': 'string'}, 'label': {'title': 'Label', 'type': 'string'}}, 'required': ['value'], 'title': 'OptionDict', 'type': 'object'}, 'TextMessage': {'properties': {'text': {'title': 'Text', 'type': 'string'}, 'type': {'const': 'text', 'default': 'text', 'title': 'Type', 'type': 'string'}}, 'required': ['text'], 'title': 'TextMessage', 'type': 'object'}}, 'items': {'$ref': '#/$defs/Message'}, 'title': 'ChatbotDataMessages', 'type': 'array', 'additional_description': None}, 'api_info_as_output': {'$defs': {'ComponentMessage': {'properties': {'component': {'title': 'Component', 'type': 'string'}, 'value': {'title': 'Value'}, 'constructor_args': {'additionalProperties': True, 'title': 'Constructor Args', 'type': 'object'}, 'props': {'additionalProperties': True, 'title': 'Props', 'type': 'object'}, 'type': {'const': 'component', 'default': 'component', 'title': 'Type', 'type': 'string'}}, 'required': ['component', 'value', 'constructor_args', 'props'], 'title': 'ComponentMessage', 'type': 'object'}, 'FileData': {'description': 'The FileData class is a subclass of the GradioModel class that represents a file object within a Gradio interface. It is used to store file data and metadata when a file is uploaded.\n\nAttributes:\n    path: The server file path where the file is stored.\n    url: The normalized server URL pointing to the file.\n    size: The size of the file in bytes.\n    orig_name: The original filename before upload.\n    mime_type: The MIME type of the file.\n    is_stream: Indicates whether the file is a stream.\n    meta: Additional metadata used internally (should not be changed).', 'properties': {'path': {'title': 'Path', 'type': 'string'}, 'url': {'anyOf': [{'type': 'string'}, {'type': 'null'}], 'default': None, 'title': 'Url'}, 'size': {'anyOf': [{'type': 'integer'}, {'type': 'null'}], 'default': None, 'title': 'Size'}, 'orig_name': {'anyOf': [{'type': 'string'}, {'type': 'null'}], 'default': None, 'title': 'Orig Name'}, 'mime_type': {'anyOf': [{'type': 'string'}, {'type': 'null'}], 'default': None, 'title': 'Mime Type'}, 'is_stream': {'default': False, 'title': 'Is Stream', 'type': 'boolean'}, 'meta': {'$ref': '#/$defs/FileDataMeta'}}, 'required': ['path'], 'title': 'FileData', 'type': 'object'}, 'FileDataMeta': {'properties': {'_type': {'const': 'gradio.FileData', 'title': 'Type', 'type': 'string'}}, 'required': ['_type'], 'title': 'FileDataMeta', 'type': 'object'}, 'FileMessage': {'properties': {'file': {'$ref': '#/$defs/FileData'}, 'alt_text': {'anyOf': [{'type': 'string'}, {'type': 'null'}], 'default': None, 'title': 'Alt Text'}, 'type': {'const': 'file', 'default': 'file', 'title': 'Type', 'type': 'string'}}, 'required': ['file'], 'title': 'FileMessage', 'type': 'object'}, 'Message': {'properties': {'role': {'title': 'Role', 'type': 'string'}, 'metadata': {'anyOf': [{'$ref': '#/$defs/MetadataDict'}, {'type': 'null'}], 'default': None}, 'content': {'items': {'anyOf': [{'$ref': '#/$defs/TextMessage'}, {'$ref': '#/$defs/FileMessage'}, {'$ref': '#/$defs/ComponentMessage'}]}, 'title': 'Content', 'type': 'array'}, 'options': {'anyOf': [{'items': {'$ref': '#/$defs/OptionDict'}, 'type': 'array'}, {'type': 'null'}], 'default': None, 'title': 'Options'}}, 'required': ['role', 'content'], 'title': 'Message', 'type': 'object'}, 'MetadataDict': {'description': 'A typed dictionary to represent metadata for a message in the Chatbot component. An\ninstance of this dictionary is used for the `metadata` field in a ChatMessage when\nthe chat message should be displayed as a thought.\nParameters:\n    title: The title of the "thought" message. Required if the message is to be displayed as a thought.\n    id: The ID of the message. Only used for nested thoughts. Nested thoughts can be nested by setting the parent_id to the id of the parent thought.\n    parent_id: The ID of the parent message. Only used for nested thoughts.\n    log: A string message to display next to the thought title in a subdued font.\n    duration: The duration of the message in seconds. Appears next to the thought title in a subdued font inside a parentheses.\n    status: if set to `"pending"`, a spinner appears next to the thought title and the accordion is initialized open.  If `status` is `"done"`, the thought accordion is initialized closed. If `status` is not provided, the thought accordion is initialized open and no spinner is displayed.', 'properties': {'title': {'title': 'Title', 'type': 'string'}, 'id': {'anyOf': [{'type': 'integer'}, {'type': 'string'}], 'title': 'Id'}, 'parent_id': {'anyOf': [{'type': 'integer'}, {'type': 'string'}], 'title': 'Parent Id'}, 'log': {'title': 'Log', 'type': 'string'}, 'duration': {'title': 'Duration', 'type': 'number'}, 'status': {'enum': ['pending', 'done'], 'title': 'Status', 'type': 'string'}}, 'title': 'MetadataDict', 'type': 'object'}, 'OptionDict': {'description': 'A typed dictionary to represent an option in a ChatMessage. A list of these\ndictionaries is used for the `options` field in a ChatMessage.\nParameters:\n    value: The value to return when the option is selected.\n    label: The text to display in the option, if different from the value.', 'properties': {'value': {'title': 'Value', 'type': 'string'}, 'label': {'title': 'Label', 'type': 'string'}}, 'required': ['value'], 'title': 'OptionDict', 'type': 'object'}, 'TextMessage': {'properties': {'text': {'title': 'Text', 'type': 'string'}, 'type': {'const': 'text', 'default': 'text', 'title': 'Type', 'type': 'string'}}, 'required': ['text'], 'title': 'TextMessage', 'type': 'object'}}, 'items': {'$ref': '#/$defs/Message'}, 'title': 'ChatbotDataMessages', 'type': 'array', 'additional_description': None}, 'example_inputs': [{'role': 'user', 'metadata': None, 'content': [{'text': 'Hello!', 'type': 'text'}], 'options': None}, {'role': 'assistant', 'metadata': None, 'content': [{'text': 'How can I help you?', 'type': 'text'}], 'options': None}]}, {'id': 30, 'type': 'row', 'props': {'variant': 'default', 'visible': True, 'elem_classes': ['chat-input-row'], 'equal_height': False, 'show_progress': False, 'preserved_by_key': [], 'name': 'row'}, 'skip_api': True, 'component_class_id': '6f8a6130c432a547a95664f7f2f2a01fd8019e8e9b05282a61688092ea105a01', 'key': None}, {'id': 31, 'type': 'textbox', 'props': {'type': 'text', 'lines': 1, 'placeholder': 'Describe the clinical case or ask a follow-up question...', 'show_label': False, 'container': False, 'scale': 8, 'min_width': 160, 'visible': True, 'autofocus': False, 'autoscroll': True, 'elem_classes': [], 'preserved_by_key': ['value'], 'rtl': False, 'buttons': [], 'submit_btn': False, 'stop_btn': False, 'name': 'textbox', '_selectable': False}, 'skip_api': False, 'component_class_id': 'de54b6fd6ce8622ae36d11c1cbd43b965ca46f0c1fe283a7cec562dcb91a3208', 'key': None, 'api_info': {'type': 'string'}, 'api_info_as_input': {'type': 'string'}, 'api_info_as_output': {'type': 'string'}, 'example_inputs': 'Hello!!'}, {'id': 32, 'type': 'button', 'props': {'value': '↻', 'variant': 'secondary', 'size': 'lg', 'link_target': '_self', 'visible': True, 'interactive': True, 'elem_classes': ['btn-clear'], 'preserved_by_key': ['value'], 'scale': 0, 'min_width': 40, 'name': 'button', '_selectable': False}, 'skip_api': True, 'component_class_id': '2417a902726d3c7f260c9a2cbe2e7a1dd2fb75f94ec4fe1ce57949f9eb9b742d', 'key': None}, {'id': 33, 'type': 'button', 'props': {'value': '↑', 'variant': 'primary', 'size': 'lg', 'link_target': '_self', 'visible': True, 'interactive': True, 'elem_classes': ['btn-send'], 'preserved_by_key': ['value'], 'scale': 0, 'min_width': 40, 'name': 'button', '_selectable': False}, 'skip_api': True, 'component_class_id': '2417a902726d3c7f260c9a2cbe2e7a1dd2fb75f94ec4fe1ce57949f9eb9b742d', 'key': None}], 'css': None, 'connect_heartbeat': False, 'js': None, 'head': None, 'title': 'OncoAgent — Clinical Triage', 'space_id': None, 'enable_queue': True, 'show_error': True, 'footer_links': [], 'is_colab': False, 'max_file_size': None, 'stylesheets': [], 'theme': None, 'protocol': 'sse_v3', 'body_css': None, 'fill_height': False, 'fill_width': False, 'theme_hash': None, 'pwa': False, 'pages': [('', 'Home', True)], 'page': {'': {'layout': {'id': 0, 'children': [{'id': 1, 'children': []}, {'id': 2, 'children': []}, {'id': 3, 'children': [{'id': 4, 'children': [{'id': 5, 'children': [{'id': 6, 'children': []}, {'id': 9, 'children': [{'id': 7}, {'id': 8}]}]}, {'id': 10, 'children': [{'id': 11, 'children': [{'id': 12, 'children': []}, {'id': 13}]}, {'id': 14, 'children': [{'id': 15, 'children': []}, {'id': 16}]}]}, {'id': 17, 'children': [{'id': 18, 'children': [{'id': 19}]}, {'id': 20, 'children': [{'id': 21}]}, {'id': 22, 'children': [{'id': 23}]}]}, {'id': 24, 'children': [{'id': 25, 'children': []}, {'id': 26}]}]}, {'id': 27, 'children': [{'id': 28, 'children': [{'id': 29}, {'id': 30, 'children': [{'id': 31}, {'id': 32}, {'id': 33}]}]}]}]}]}, 'components': [1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19, 20, 21, 22, 23, 24, 25, 26, 27, 28, 29, 30, 31, 32, 33], 'dependencies': [0, 1, 2, 3]}}, 'mcp_server': False, 'i18n_translations': None, 'dependencies': [{'id': 0, 'targets': [(33, 'click')], 'inputs': [29, 31, 7, 8], 'outputs': [29, 31, 13, 16, 19, 21, 23, 26], 'backend_fn': True, 'js': None, 'queue': True, 'api_name': 'process_and_stream', 'api_description': None, 'scroll_to_output': False, 'show_progress': 'full', 'show_progress_on': None, 'batch': False, 'max_batch_size': 4, 'cancels': [], 'types': {'generator': True, 'cancel': False}, 'collects_event_data': False, 'trigger_after': None, 'trigger_only_on_success': False, 'trigger_only_on_failure': False, 'trigger_mode': 'once', 'api_visibility': 'public', 'rendered_in': None, 'render_id': None, 'connection': 'sse', 'time_limit': None, 'stream_every': 0.5, 'event_specific_args': None, 'component_prop_inputs': [], 'js_implementation': None}, {'id': 1, 'targets': [(31, 'submit')], 'inputs': [29, 31, 7, 8], 'outputs': [29, 31, 13, 16, 19, 21, 23, 26], 'backend_fn': True, 'js': None, 'queue': True, 'api_name': 'process_and_stream_1', 'api_description': None, 'scroll_to_output': False, 'show_progress': 'full', 'show_progress_on': None, 'batch': False, 'max_batch_size': 4, 'cancels': [], 'types': {'generator': True, 'cancel': False}, 'collects_event_data': False, 'trigger_after': None, 'trigger_only_on_success': False, 'trigger_only_on_failure': False, 'trigger_mode': 'once', 'api_visibility': 'public', 'rendered_in': None, 'render_id': None, 'connection': 'sse', 'time_limit': None, 'stream_every': 0.5, 'event_specific_args': None, 'component_prop_inputs': [], 'js_implementation': None}, {'id': 2, 'targets': [(32, 'click')], 'inputs': [], 'outputs': [29, 31, 7, 8, 13, 16, 19, 21, 23, 26], 'backend_fn': True, 'js': None, 'queue': True, 'api_name': 'lambda', 'api_description': None, 'scroll_to_output': False, 'show_progress': 'full', 'show_progress_on': None, 'batch': False, 'max_batch_size': 4, 'cancels': [], 'types': {'generator': False, 'cancel': False}, 'collects_event_data': False, 'trigger_after': None, 'trigger_only_on_success': False, 'trigger_only_on_failure': False, 'trigger_mode': 'once', 'api_visibility': 'public', 'rendered_in': None, 'render_id': None, 'connection': 'sse', 'time_limit': None, 'stream_every': 0.5, 'event_specific_args': None, 'component_prop_inputs': [], 'js_implementation': None}, {'id': 3, 'targets': [(None, 'load')], 'inputs': [], 'outputs': [7], 'backend_fn': True, 'js': None, 'queue': True, 'api_name': 'generate_patient_id', 'api_description': None, 'scroll_to_output': False, 'show_progress': 'full', 'show_progress_on': None, 'batch': False, 'max_batch_size': 4, 'cancels': [], 'types': {'generator': False, 'cancel': False}, 'collects_event_data': False, 'trigger_after': None, 'trigger_only_on_success': False, 'trigger_only_on_failure': False, 'trigger_mode': 'once', 'api_visibility': 'public', 'rendered_in': None, 'render_id': None, 'connection': 'sse', 'time_limit': None, 'stream_every': 0.5, 'event_specific_args': None, 'component_prop_inputs': [], 'js_implementation': None}], 'layout': {'id': 0, 'children': [{'id': 1, 'children': []}, {'id': 2, 'children': []}, {'id': 3, 'children': [{'id': 4, 'children': [{'id': 5, 'children': [{'id': 6, 'children': []}, {'id': 9, 'children': [{'id': 7}, {'id': 8}]}]}, {'id': 10, 'children': [{'id': 11, 'children': [{'id': 12, 'children': []}, {'id': 13}]}, {'id': 14, 'children': [{'id': 15, 'children': []}, {'id': 16}]}]}, {'id': 17, 'children': [{'id': 18, 'children': [{'id': 19}]}, {'id': 20, 'children': [{'id': 21}]}, {'id': 22, 'children': [{'id': 23}]}]}, {'id': 24, 'children': [{'id': 25, 'children': []}, {'id': 26}]}]}, {'id': 27, 'children': [{'id': 28, 'children': [{'id': 29}, {'id': 30, 'children': [{'id': 31}, {'id': 32}, {'id': 33}]}]}]}]}]}}

data/clinical_guides/esmo/ESMO_PMC10416694_Developing_a_core_set_of_patient_reported_outcomes.pdf

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:77ebd06e0a0c662b04910819cfb518b2ff10d5f4d4b578177476485e42515945
+size 157590

data/clinical_guides/esmo/ESMO_PMC10664856_ESMO_ASCO_Recommendations_for_a_Global_Curriculum_.pdf

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:66c17f6d3a5409f19d8f9811b0f37b83b7c667fe171ea5ee35ab40b5fc522d24
+size 912702

data/clinical_guides/esmo/ESMO_PMC10774906_ESMO_ASCO_Recommendations_for_a_Global_Curriculum_.pdf

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:951a1a53bb4b50797f5bb156acb7af7c94b4aec9cb2ee31284a787f84eae026e
+size 1125014

data/clinical_guides/esmo/ESMO_PMC11574484_Effects_of_Baduanjin_exercise_on_cancer_related_fa.pdf

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:79093a78f182b14dcdd8b9c8f7ef04fd20e92ae57c3fb4b8ef93d3279db2c623
+size 1308447

data/clinical_guides/esmo/ESMO_PMC11628549_Research_hotspots_and_trends_in_immunotherapy_for_.pdf

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:872d14742eab67fa0b0e57c6d5371738f090a459a66832eb3cad8bfba12a33ab
+size 11121286

data/clinical_guides/esmo/ESMO_PMC11662070_Characterization_of_shared_neoantigens_landscape_i.pdf

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:9a429d64c1ede46c3a5c948a72fcc9b0104aff988f1b130ed7e332732f814e2d
+size 1661268

data/clinical_guides/esmo/ESMO_PMC11856526_Multiomics_in_silico_analysis_identifies_TM4SF4_as.pdf

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:5a5e7d0d1b2834c0b33c56a48655cb3dd431fad755b81bfc400e889df76ed95f
+size 4129503

data/clinical_guides/esmo/ESMO_PMC12218492_Plain_language_summary_of_the_THOR_Cohort_1_study_.pdf

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:66bbb6fd168c8851ec615b28eda7bfd62ab1b1d3c48eb69762a3c7cdcd071da7
+size 1964985

data/clinical_guides/esmo/ESMO_PMC12306965_Systematic_critical_appraisal_of_GRADE_recommendat.pdf

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:d11f20a1198835b1d7c6f28a0284fd5a904722d50a57e760bef7ce7e2008b89e
+size 199162

data/clinical_guides/esmo/ESMO_PMC12381471_Prevention_and_treatment_of_venous_thromboembolism.pdf

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:ffb92f9ce4d44554acfc38d0e341e2a2f93cbf961160305cc04aab149097e732
+size 431486

data/clinical_guides/esmo/ESMO_PMC12733557_How_to_Read_a_Next_Generation_Sequencing_Report_fo.pdf

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:d6ee5890e0fb3c71537398db23428e6284a671a35d1e4beaad945019e9ad3e99
+size 2600162

data/clinical_guides/esmo/ESMO_PMC12836719_Adoption_of_electronic_patient_reported_outcomes_i.pdf

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:42fbd736ea71aa5767847068bafc22c04cfa65cf4ffba86bafe6e28a38b14d7e
+size 521364

data/clinical_guides/esmo/ESMO_PMC12925129_Development_and_formative_evaluation_of_a_follow_u.pdf

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:da3cf1fd73a6ba56c436ddb96011fe91ca21825b4dc1dd8a3605e8a101acc702
+size 663158

data/clinical_guides/esmo/ESMO_PMC12982907_Cost_utility_Analysis_of_R_CHOP_vs_CHOP_in_Patient.pdf

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:aa28cae3e9bd3d1a0ee96671f77007ac24815cce004032daa84efb1e7cfcc984
+size 6726400

data/clinical_guides/esmo/ESMO_PMC7617288_Randomised_Trial_of_No__Short_term__or_Long_term_A.pdf

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:ee8b7a0ac8377222884d97192021783710b5875caeadadd074ad2b268ad35d82
+size 695629

data/clinical_guides/esmo/ESMO_PMC8267298_An_evaluation_of_the_reporting_quality_in_clinical.pdf

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:1403746a915c1357ff17c694f4a40ff197e28d8b6ac07ab29b5d6b74d4ff6308
+size 358975

data/clinical_guides/nccn/Patient guidelines/Guidelines for Detection, Prevention, & Risk Reduction/breastcancerscreening-patient.pdf

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:60ab6f0176fc3e6804765f489bb521bce0249023eed0b4c8124da6d0daf3c676
+size 3082254

data/clinical_guides/nccn/Patient guidelines/Guidelines for Detection, Prevention, & Risk Reduction/colorectal-screening-patient.pdf

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:b5a77a23f2329bee0e1859b97c74fa77a8b5ea3d16dba101b511b451d326f915
+size 1778124

data/clinical_guides/nccn/Patient guidelines/Guidelines for Detection, Prevention, & Risk Reduction/genetics-patient.pdf

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:a0426fa43388588f40abeb4cc6f9768d73cf410f6615f5f1678fe88d2e0fb791
+size 3067435

data/clinical_guides/nccn/Patient guidelines/Guidelines for Detection, Prevention, & Risk Reduction/lung_screening-patient.pdf

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:e9348bdb68b6336f80355adfdbda0ea593e9e257034c8844370aad5db891be39
+size 1909020

data/clinical_guides/nccn/Patient guidelines/Guidelines for Detection, Prevention, & Risk Reduction/prostate-screening-patient.pdf

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:8479e1d957db2fd12aa389ae994151d30f6e2ab6d5e1bb0d6628b464f640625b
+size 1225976

data/clinical_guides/nccn/Patient guidelines/Guidelines for Specific Populations/aya-patient.pdf

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:44777805ee2d1ca752276bd8d1ebeb9cf6da92d57b65e4af2ac410493d68f4f9
+size 1929361

data/clinical_guides/nccn/Patient guidelines/Guidelines for Supportive Care/GVDH-patient-guideline.pdf

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:7dcc8e5f23b17f9fb74bc8c710fc50b49756bd581cf4cc2168e6ad338034439a
+size 1352636

data/clinical_guides/nccn/Patient guidelines/Guidelines for Supportive Care/bloodclots-patient.pdf

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:3cb4e08c8e6e7264b06abb7508a602764d46c034f42c4213e71f1978ce984334
+size 3144306

data/clinical_guides/nccn/Patient guidelines/Guidelines for Supportive Care/distress-patient.pdf

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:69865878db38ad4692274187bce987bf8520f102a0023cde398d4cf423e76285
+size 694090

data/clinical_guides/nccn/Patient guidelines/Guidelines for Supportive Care/fatigue-patient.pdf

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:44016e3c5b2bdfccf44d43788c726df000fa4062e81ee5c079890c635b47ac7f
+size 696543

data/clinical_guides/nccn/Patient guidelines/Guidelines for Supportive Care/immunotherapy-checkpoint-patient.pdf

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:713cffcdf1ffa176f38081b61addabb6963a24be1fa675f1d1233589978661b3
+size 1789367

data/clinical_guides/nccn/Patient guidelines/Guidelines for Supportive Care/immunotherapy-se-car-tcell-patient.pdf

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:288f6c80406f8a3eee289676ca4a4eb84d48b736dfd6fafd06c54f025586178d
+size 1847930

data/clinical_guides/nccn/Patient guidelines/Guidelines for Supportive Care/low-blood-patient.pdf

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:ab9345cff6ed2c10fbd5e60f87228c579c8ddcb922c8f4151f3bbf41ff51ab6e
+size 911272

data/clinical_guides/nccn/Patient guidelines/Guidelines for Supportive Care/nausea-patient.pdf

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:8287abbd7b9f7d5c6acff176aa846711fd9558e23b9e238324dc77995069036a
+size 3485802

data/clinical_guides/nccn/Patient guidelines/Guidelines for Supportive Care/palliative-patient.pdf

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:5834615b9e5ac24a860d79c469757a1ea4490db788d93bb98535f424f4daab4a
+size 1256795

data/clinical_guides/nccn/Patient guidelines/Guidelines for Supportive Care/quitting-smoking-patient.pdf

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:9793461e32b296e0d095e05e646aa1c5c4652ef7591f0f83433fe7b0d1cdabaf
+size 1568449