File size: 3,311 Bytes
4745039 d0b7211 2f495b0 904dfee d0b7211 4745039 d0b7211 2f495b0 904dfee d0b7211 2f495b0 d0b7211 904dfee d0b7211 2f495b0 d0b7211 4745039 d0b7211 4745039 d0b7211 4745039 d0b7211 4745039 d0b7211 4745039 d0b7211 4745039 d0b7211 4745039 | 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 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 | # ARIA Project Status
_Last updated: May 5, 2026_
## What Was Built
Multi-agent pipeline in agent/agent.py with 5 stages:
1. Query Architect: generates 5 MeSH-optimised PubMed queries via Qwen2.5-72B on AMD MI300X
2. Literature Scout: fetches from PubMed and Europe PMC in parallel, deduplicates by PMID
3. PRISMA Filter: automatic inclusion/exclusion screening with one-line reasons, user can override any decision
4. Evidence Synthesiser: structured synthesis with Background, Key Findings, Level of Evidence, Conflicting Evidence, Research Gaps, Clinical Implications
5. Citation Builder: formatted references with PMID, synthesis runs on PRISMA-included papers only
## Additional Features
* SSE streaming: real-time 5-stage progress bar with percentage
* PRISMA-style paper screening: automated include/exclude with rationale, user override buttons
* Evidence confidence scoring: green/yellow/red badges on each synthesis section with hover tooltips
* Abstract viewer: click any citation to expand full abstract inline
* PDF export: download formatted report via ReportLab
* Selective literature review: checkboxes on citations, user picks papers, generates focused academic paragraph
* Predictive model: constructive and destructive forecasts as a post-synthesis stage
* Evidence comparison table: LLM extracts structured data table with real author names, no duplicates
* Follow-up Q&A: ask follow-up questions after synthesis, answered using already-fetched papers
* Query refinement suggestions: 3 AI-generated follow-up research questions based on synthesis gaps
* Session history: queries saved to sessions.json, reloadable from sidebar
* Rate limit retry logic: automatic backoff on API errors
* SSL patch for PubMed and Europe PMC Entrez on corporate/university networks
## Tech Stack
* LLM: Qwen2.5-72B-Instruct on AMD MI300X via vLLM 0.17.1
* Agent Framework: LangGraph + LangChain
* Literature Retrieval: BioPython Entrez / PubMed NCBI + Europe PMC
* Web Framework: Flask with SSE streaming
* PDF: ReportLab
* Frontend: HTML, CSS, vanilla JS
* Runtime: Windows 11, Python 3.11, RTX 3060 12GB (local) + AMD MI300X 192GB (inference)
## API Endpoints
* GET / � main UI
* GET /stream � SSE pipeline stream
* POST /query � standard pipeline (fallback)
* POST /score � confidence scoring
* POST /predict � predictive model
* POST /selective-review � focused literature review from selected papers
* POST /extract-table � evidence comparison table
* POST /export-pdf � PDF report download
* POST /followup � follow-up question against existing synthesis
* POST /suggest-queries � 3 AI-generated follow-up research questions
* GET /sessions � load query history
* POST /sessions/save � save completed query
## Environment
* VLLM_BASE_URL: set via HF Space environment variable
* VLLM_API_KEY: EMPTY
* Python venv at ./venv
* Start server: venv\Scripts\activate && python app.py
* AMD MI300X: DigitalOcean droplet, 192GB VRAM
## Deployment
* HF Space: https://huggingface.co/spaces/lablab-ai-amd-developer-hackathon/glitch-squad-biomedical-assistant
* GitHub: https://github.com/azlaan428/glitch-squad-biomedical-assistant
## What Remains
1. Demo video (under 5 minutes)
2. Final lablab.ai submission
|