app.py

"""
Sakhi (सखी) — ASHA Health Worker AI Companion
================================================
Hindi voice → structured MCTS/HMIS forms + danger sign detection
powered by Gemma 4 E4B (fine-tuned via Unsloth).

This module is the pipeline library (ASR + extraction + validation). The
React UI is served by api.py; this file is not run directly.
"""
import os
import re
import json
import time

os.environ["TORCH_COMPILE_DISABLE"] = "1"
os.environ["TORCHDYNAMO_DISABLE"] = "1"

# ============================================================
# CONFIGURATION
# ============================================================
MODEL_PATH = "./models/checkpoints/final"
MAX_SEQ_LENGTH = 4096

# Ollama config — set OLLAMA_MODEL to use Ollama instead of Unsloth
# Use "sakhi" once fine-tuned GGUF is registered, or base model for now
OLLAMA_MODEL = os.environ.get("OLLAMA_MODEL", "gemma4:e4b-it-q4_K_M")
USE_OLLAMA = os.environ.get("USE_OLLAMA", "1") == "1"
USE_FUNCTION_CALLING = os.environ.get("USE_FUNCTION_CALLING", "1") == "1"

# Whisper config. Default = the CTranslate2-converted mirror of collabora's
# Hindi fine-tune of whisper-large-v2 (selected after session 19's real-voice
# validation pass). faster-whisper requires CT2 format; the original
# collabora/ repo is transformers format and won't load directly.
# Override with WHISPER_MODEL for evals against other variants. Local dev
# with a pre-converted CT2 directory at models/whisper-hindi-ct2/ takes
# precedence over this env var — see warm_whisper().
WHISPER_MODEL = os.environ.get("WHISPER_MODEL", "Tushar9802/whisper-large-v2-hindi-ct2")

# System prompts (same as training)
FORM_SYSTEM_PROMPT = (
    "You are a clinical data extraction system for India's ASHA health worker program. "
    "Extract structured data from the Hindi/Hinglish home visit conversation into the requested JSON schema. "
    "ONLY extract information explicitly stated in the conversation. Use null for any field not mentioned.\n\n"
    "STRICT RULES:\n"
    "1. Do NOT invent names, dates, phone numbers, or addresses. If the patient is only called 'दीदी' or 'बहन', set name to null.\n"
    "2. If age is not explicitly stated as a number, set age to null. Do NOT guess from context.\n"
    "3. If blood group, HIV status, or other lab tests are not discussed, they MUST be null — never assume 'negative' or a default group.\n"
    "4. If the conversation has no speaker labels (ASHA/Patient), still extract data but be extra strict about nulls.\n"
    "5. Numbers may appear as Hindi words (e.g., 'एक सो दस बटा सत्तर' = 110/70). Convert them to digits.\n"
    "Return valid JSON only."
)

DANGER_SYSTEM_PROMPT = (
    "You are a clinical danger sign detection system for India's ASHA health worker program. "
    "Analyze the Hindi/Hinglish home visit conversation for NHM-defined danger signs.\n\n"
    "STRICT RULES:\n"
    "1. ONLY flag a danger sign if the EXACT words proving it appear in the conversation.\n"
    "2. utterance_evidence MUST be a verbatim copy-paste from the conversation — do NOT paraphrase or fabricate.\n"
    "3. If a vital sign is NORMAL (e.g., BP 110/70, temperature 37°C), that is NOT a danger sign.\n"
    "4. Most routine visits have ZERO danger signs. Return an empty danger_signs array when none exist.\n"
    "5. When in doubt, do NOT flag — a missed flag is better than a false alarm.\n"
    "Return valid JSON only."
)

# ============================================================
# EXAMPLE TRANSCRIPTS (for demo)
# ============================================================
EXAMPLE_TRANSCRIPTS = [
    [
        "ANC Visit — Normal",
        (
            "ASHA: नमस्ते, कैसे हैं आप?\n"
            "Patient: नमस्ते दीदी, मैं ठीक हूँ।\n"
            "ASHA: अच्छा है। मैं आपका चेकअप करने आई हूँ। चलिए, पहले आपका BP चेक कर लेती हूँ।\n"
            "Patient: ठीक है।\n"
            "ASHA: आपका BP 110/70 है, बिल्कुल ठीक है। अब वजन देखती हूँ... 58 kg है। पिछली बार 56 था, तो अच्छा बढ़ रहा है।\n"
            "Patient: हाँ, मैं अच्छा खा रही हूँ।\n"
            "ASHA: बहुत अच्छा! Hb कितना आया था पिछली बार?\n"
            "Patient: डॉक्टर ने कहा था 11.5 है।\n"
            "ASHA: ये तो बहुत अच्छा है। IFA की गोलियाँ ले रही हैं?\n"
            "Patient: हाँ, रोज़ लेती हूँ।\n"
            "ASHA: TT का टीका लगा?\n"
            "Patient: हाँ, पहला लग गया है।\n"
            "ASHA: बच्चे की हलचल कैसी है?\n"
            "Patient: बहुत हिलता-डुलता है, ठीक है।\n"
            "ASHA: बहुत अच्छा। आप लगभग 24 हफ्ते की हैं। डिलीवरी के लिए कहाँ जाएँगी?\n"
            "Patient: PHC में।\n"
            "ASHA: गाड़ी का इंतज़ाम है?\n"
            "Patient: हाँ, पति की गाड़ी है।\n"
            "ASHA: ठीक है। अगली बार 2 हफ्ते बाद आऊँगी। कोई तकलीफ़ हो तो फ़ोन कर दीजिए।\n"
            "Patient: ठीक है दीदी, धन्यवाद।"
        ),
    ],
    [
        "ANC Visit — Preeclampsia (DANGER)",
        (
            "ASHA: नमस्ते दीदी, कैसे हैं?\n"
            "Patient: दीदी, मुझे बहुत सिरदर्द हो रहा है कल से।\n"
            "ASHA: अच्छा, और कोई तकलीफ़?\n"
            "Patient: हाँ, आँखों के सामने धुंधला दिखता है कभी-कभी। और चेहरे पर सूजन भी आ गई है।\n"
            "ASHA: ये तो ठीक नहीं है। मैं BP चेक करती हूँ... आपका BP 155/100 आ रहा है। ये बहुत ज़्यादा है।\n"
            "Patient: क्या करें दीदी?\n"
            "ASHA: आपको तुरंत PHC जाना होगा। ये गंभीर हो सकता है। आप कितने महीने की हैं?\n"
            "Patient: लगभग 8 महीने।\n"
            "ASHA: पैरों में सूजन है?\n"
            "Patient: हाँ, काफी सूजन है।\n"
            "ASHA: मैं अभी गाड़ी का इंतज़ाम करती हूँ। आपको आज ही PHC ले चलती हूँ।"
        ),
    ],
    [
        "PNC — Newborn not feeding (DANGER)",
        (
            "ASHA: नमस्ते, कैसे हैं? बच्चा कैसा है?\n"
            "Mother: दीदी, बच्चा बहुत सोता रहता है। दूध भी ठीक से नहीं पीता।\n"
            "ASHA: कब से ऐसा है?\n"
            "Mother: कल से। पहले ठीक था, अब लगभग 12 घंटे से दूध नहीं पिया।\n"
            "ASHA: बच्चे का रोना कैसा है?\n"
            "Mother: बहुत कमज़ोर आवाज़ में रोता है।\n"
            "ASHA: तापमान चेक करती हूँ... 100.5 डिग्री है। बुखार है। और बच्चा सुस्त लग रहा है।\n"
            "Mother: क्या करें?\n"
            "ASHA: ये IMNCI के danger signs हैं। बच्चे को तुरंत PHC ले जाना होगा। मैं गाड़ी बुलाती हूँ।"
        ),
    ],
    [
        "Child Health — Routine visit",
        (
            "ASHA: नमस्ते, बच्चा कैसा है?\n"
            "Mother: बिल्कुल ठीक है दीदी। खूब खाता है, खेलता है।\n"
            "ASHA: बहुत अच्छा! वजन देखती हूँ... 8.5 kg है। 9 महीने के लिए अच्छा है।\n"
            "Mother: हाँ, दाल-चावल, केला सब खाता है अब।\n"
            "ASHA: Vitamin A की दवाई दी थी पिछली बार?\n"
            "Mother: हाँ, 6 महीने में दी थी।\n"
            "ASHA: अच्छा। अब deworming भी देनी है। और टीके सब लगे हैं?\n"
            "Mother: हाँ, सब समय पर लगे हैं।\n"
            "ASHA: बहुत अच्छा। बच्चा बैठता है, घुटनों पर चलता है?\n"
            "Mother: हाँ, सब करता है। बोलने भी लगा है थोड़ा।\n"
            "ASHA: बढ़िया है। अगली बार 3 महीने बाद आऊँगी।"
        ),
    ],
]


# ============================================================
# SCHEMA LOADING
# ============================================================
def load_schema(name):
    with open(f"configs/schemas/{name}.json", "r", encoding="utf-8") as f:
        return json.load(f)


SCHEMAS = {}
VISIT_TYPE_MAP = {
    "anc_visit": "anc_visit",
    "pnc_visit": "pnc_visit",
    "delivery": "delivery",
    "child_health": "child_health",
}


def init_schemas():
    global SCHEMAS
    for name in ["anc_visit", "pnc_visit", "delivery", "child_health", "danger_signs"]:
        SCHEMAS[name] = load_schema(name)


# ============================================================
# MODEL LOADING
# ============================================================
_model = None
_tokenizer = None


def load_model():
    global _model, _tokenizer
    if _model is not None:
        return _model, _tokenizer

    import torch
    torch._dynamo.config.suppress_errors = True
    from unsloth import FastLanguageModel

    print("[MODEL] Loading Gemma 4 E4B fine-tuned model...")
    _model, _tokenizer = FastLanguageModel.from_pretrained(
        model_name=MODEL_PATH,
        max_seq_length=MAX_SEQ_LENGTH,
        load_in_4bit=True,
    )
    FastLanguageModel.for_inference(_model)
    print("[MODEL] Model loaded.")
    return _model, _tokenizer


# ============================================================
# TRANSCRIPT POST-PROCESSING (delegated to src/hindi_normalize)
# ============================================================
from src.hindi_normalize import normalize_transcript as postprocess_transcript


_whisper_model = None

def warm_whisper():
    """Eagerly load the Whisper model into VRAM. Idempotent — safe to call
    multiple times; subsequent calls return the cached singleton. Called from
    FastAPI's startup hook so the first user audio request lands hot."""
    global _whisper_model
    if _whisper_model is not None:
        return _whisper_model
    from faster_whisper import WhisperModel
    ct2_path = os.path.join(os.path.dirname(__file__), "models", "whisper-hindi-ct2")
    if os.path.exists(ct2_path):
        print(f"[ASR] Loading CTranslate2 model from {ct2_path}...")
        _whisper_model = WhisperModel(ct2_path, device="cuda", compute_type="float16")
    else:
        print(f"[ASR] Loading {WHISPER_MODEL} from HuggingFace Hub...")
        _whisper_model = WhisperModel(WHISPER_MODEL, device="cuda", compute_type="float16")
    print("[ASR] Whisper loaded.")
    return _whisper_model


def transcribe_audio(audio_path):
    """Transcribe audio using the configured Whisper model via faster-whisper (CTranslate2)."""
    warm_whisper()

    print("[ASR] Transcribing...")
    segments, info = _whisper_model.transcribe(
        audio_path,
        language="hi",
        task="transcribe",
        vad_filter=True,
        beam_size=1,
        temperature=0.0,
        condition_on_previous_text=False,
    )
    transcript = " ".join(seg.text.strip() for seg in segments)

    transcript = postprocess_transcript(transcript)

    print(f"[ASR] Transcript ({len(transcript)} chars)")
    return transcript


def run_inference(system_prompt, user_prompt):
    """Run model inference via Ollama or Unsloth, return parsed JSON or raw text."""
    if USE_OLLAMA:
        return _run_inference_ollama(system_prompt, user_prompt)
    return _run_inference_unsloth(system_prompt, user_prompt)


def translate_to_english(hindi_text):
    """Translate Hindi / Hinglish home-visit text to English via the same
    Gemma model already loaded in VRAM. On-demand only — never on the
    main extraction path. Returns plain English text (not JSON)."""
    import ollama

    text = (hindi_text or "").strip()
    if not text:
        return ""

    t0 = time.time()
    resp = ollama.chat(
        model=OLLAMA_MODEL,
        messages=[
            {"role": "system", "content": (
                "Translate the following Hindi or Hinglish conversation into clear, natural English. "
                "Preserve speaker labels (ASHA / Patient / Mother) and clinical numbers exactly. "
                "Do not add commentary or explanations — output ONLY the translation."
            )},
            {"role": "user", "content": text},
        ],
        options={"temperature": 0.1, "num_ctx": 4096, "num_gpu": 999},
        keep_alive=os.environ.get("OLLAMA_KEEP_ALIVE", "10m"),
    )
    elapsed = time.time() - t0

    out = resp.message.content.strip()
    tok_s = resp.eval_count / (resp.eval_duration / 1e9) if resp.eval_duration else 0
    print(f"[LLM] Translate: {elapsed:.1f}s ({resp.eval_count} tok, {tok_s:.0f} tok/s)")
    return out


def _run_inference_ollama(system_prompt, user_prompt):
    """Run inference via Ollama API — fast GGUF on GPU with JSON mode."""
    import ollama

    t0 = time.time()
    resp = ollama.chat(
        model=OLLAMA_MODEL,
        messages=[
            {"role": "system", "content": system_prompt},
            {"role": "user", "content": user_prompt},
        ],
        format="json",
        options={"temperature": 0.1, "num_ctx": 4096, "num_gpu": 999},
        keep_alive=os.environ.get("OLLAMA_KEEP_ALIVE", "10m"),
    )
    elapsed = time.time() - t0

    response = resp.message.content
    tok_s = resp.eval_count / (resp.eval_duration / 1e9) if resp.eval_duration else 0
    print(f"[LLM] Ollama: {elapsed:.1f}s ({resp.eval_count} tok, {tok_s:.0f} tok/s)")

    # format="json" guarantees valid JSON — parse directly
    try:
        parsed = json.loads(response)
    except json.JSONDecodeError:
        print(f"[WARN] Ollama JSON mode parse failed, falling back to heuristic parser")
        parsed = _parse_json_response(response)
    return {"raw": response, "parsed": parsed, "time_s": elapsed}


# ============================================================
# FUNCTION CALLING — Gemma 4 native tool use
# ============================================================

def _build_form_tool(visit_type):
    """Build extract_form tool definition from the visit's JSON schema."""
    schema_key = VISIT_TYPE_MAP.get(visit_type, "anc_visit")
    schema = SCHEMAS.get(schema_key, SCHEMAS["anc_visit"])
    return {
        "type": "function",
        "function": {
            "name": "extract_form",
            "description": (
                f"Extract structured {schema_key.replace('_', ' ')} form data from the "
                "ASHA home visit conversation. ONLY extract information explicitly stated. "
                "Use null for any field not mentioned."
            ),
            "parameters": schema,
        },
    }


TOOL_FLAG_DANGER_SIGN = {
    "type": "function",
    "function": {
        "name": "flag_danger_sign",
        "description": (
            "Flag a single danger sign detected in the patient conversation. "
            "Call once per danger sign found. Do NOT call if no danger signs exist. "
            "The evidence field MUST be an exact verbatim quote from the conversation."
        ),
        "parameters": {
            "type": "object",
            "properties": {
                "sign": {
                    "type": "string",
                    "description": "Standard NHM danger sign name (e.g., severe_preeclampsia, severe_anemia)",
                },
                "category": {
                    "type": "string",
                    "enum": ["immediate_referral", "urgent_care", "monitor_closely"],
                },
                "clinical_value": {
                    "type": ["string", "null"],
                    "description": "Measured value if applicable (e.g., '145/95', '38.5C')",
                },
                "utterance_evidence": {
                    "type": "string",
                    "description": "REQUIRED: exact verbatim quote from conversation proving this sign",
                },
            },
            "required": ["sign", "category", "utterance_evidence"],
        },
    },
}

TOOL_ISSUE_REFERRAL = {
    "type": "function",
    "function": {
        "name": "issue_referral",
        "description": (
            "Issue a referral decision based on detected danger signs. "
            "Only call if danger signs warrant referral. Do NOT call for routine visits."
        ),
        "parameters": {
            "type": "object",
            "properties": {
                "urgency": {
                    "type": "string",
                    "enum": ["immediate", "within_24h", "routine"],
                },
                "facility": {
                    "type": ["string", "null"],
                    "enum": ["PHC", "CHC", "district_hospital", "FRU", None],
                },
                "reason": {
                    "type": "string",
                    "description": "Brief clinical reasoning for referral",
                },
            },
            "required": ["urgency", "facility", "reason"],
        },
    },
}

DANGER_FC_SYSTEM_PROMPT = (
    "You are a clinical danger sign detection system for India's ASHA health worker program.\n\n"
    "Analyze the conversation and use the provided tools:\n"
    "1. flag_danger_sign — call ONCE per danger sign found. Evidence MUST be a verbatim quote from the conversation. "
    "If NO danger signs exist, do NOT call any tool.\n"
    "2. issue_referral — call only if danger signs warrant referral to a facility.\n\n"
    "STRICT RULES:\n"
    "- ONLY flag a danger sign if the EXACT words proving it appear in the conversation.\n"
    "- utterance_evidence MUST be a verbatim copy-paste from the conversation — do NOT paraphrase.\n"
    "- If a vital sign is NORMAL (e.g., BP 110/70, temperature 37°C), that is NOT a danger sign.\n"
    "- Most routine visits have ZERO danger signs. Do NOT call any tools for normal visits.\n"
    "- When in doubt, do NOT flag — a missed flag is better than a false alarm."
)


def _run_danger_fc(transcript, visit_type):
    """Run danger sign detection via function calling (flag_danger_sign + issue_referral tools)."""
    import ollama

    tools = [TOOL_FLAG_DANGER_SIGN, TOOL_ISSUE_REFERRAL]

    t0 = time.time()
    resp = ollama.chat(
        model=OLLAMA_MODEL,
        messages=[
            {"role": "system", "content": DANGER_FC_SYSTEM_PROMPT},
            {"role": "user", "content": (
                f"Analyze this ASHA home visit conversation for danger signs.\n\n"
                f"Visit type: {visit_type}\n\n"
                f"{transcript}"
            )},
        ],
        tools=tools,
        options={"temperature": 0.1, "num_ctx": 4096, "num_gpu": 999},
        keep_alive=os.environ.get("OLLAMA_KEEP_ALIVE", "10m"),
    )
    elapsed = time.time() - t0

    tok_s = resp.eval_count / (resp.eval_duration / 1e9) if resp.eval_duration else 0
    print(f"[LLM] Danger FC: {elapsed:.1f}s ({resp.eval_count} tok, {tok_s:.0f} tok/s)")

    danger_signs = []
    referral = None
    tool_calls_raw = []

    if resp.message.tool_calls:
        for tc in resp.message.tool_calls:
            fname = tc.function.name
            args = tc.function.arguments
            tool_calls_raw.append({"function": fname, "arguments": args})

            if fname == "flag_danger_sign":
                danger_signs.append(args)
            elif fname == "issue_referral":
                referral = args

        print(f"[LLM] Tool calls: {len(resp.message.tool_calls)} "
              f"(danger_signs={len(danger_signs)}, "
              f"referral={'yes' if referral else 'no'})")
    else:
        print(f"[LLM] No tool calls — no danger signs detected")

    return {
        "danger_signs": danger_signs,
        "referral": referral,
        "tool_calls": tool_calls_raw,
        "time_s": elapsed,
    }


def _normalize_fc_form(raw, visit_type):
    """Normalize function calling form output to match the expected schema structure.

    The model sometimes uses free-form keys (blood_pressure: "110/70") instead
    of schema keys (bp_systolic: 110, bp_diastolic: 70), or nests data
    differently. This flattens and remaps to the canonical form.
    """
    if not raw or not isinstance(raw, dict):
        return raw

    # Recursively collect all key-value pairs from the raw output
    def _collect(d, prefix=""):
        items = {}
        if isinstance(d, dict):
            for k, v in d.items():
                key = f"{prefix}.{k}" if prefix else k
                if isinstance(v, dict):
                    items.update(_collect(v, key))
                else:
                    items[key] = v
                    # Also store under the leaf key for simple matching
                    items[k] = v
        return items

    flat = _collect(raw)

    # Build a clean output matching schema structure
    schema_key = VISIT_TYPE_MAP.get(visit_type, "anc_visit")
    schema = SCHEMAS.get(schema_key, SCHEMAS.get("anc_visit", {}))
    result = {}

    # Walk schema top-level sections and fill from flat values
    for section_name, section_def in schema.get("properties", {}).items():
        if section_def.get("type") == "object":
            section_data = {}
            for field_name in section_def.get("properties", {}).keys():
                # Try exact match first, then look through flat keys
                val = flat.get(f"{section_name}.{field_name}") or flat.get(field_name)
                if val is not None:
                    section_data[field_name] = val
            if section_data:
                result[section_name] = section_data
        elif section_def.get("type") == "array":
            val = flat.get(section_name)
            if isinstance(val, list):
                result[section_name] = val
            else:
                result[section_name] = []
        else:
            val = flat.get(section_name)
            if val is not None:
                result[section_name] = val

    # ── BP splitting: "110/70" → bp_systolic=110, bp_diastolic=70 ──
    vitals = result.get("vitals", {})
    bp_raw = flat.get("blood_pressure") or flat.get("bp") or flat.get("vitals.blood_pressure")
    if bp_raw and isinstance(bp_raw, str) and "/" in bp_raw:
        parts = bp_raw.split("/")
        try:
            if "bp_systolic" not in vitals or vitals.get("bp_systolic") is None:
                vitals["bp_systolic"] = int(parts[0].strip())
            if "bp_diastolic" not in vitals or vitals.get("bp_diastolic") is None:
                vitals["bp_diastolic"] = int(parts[1].strip())
        except (ValueError, IndexError):
            pass

    # ── Infant/child weight normalization (before vitals, to avoid misplacement) ──
    # PNC: infant_assessment.weight_kg, Delivery: infant.birth_weight_kg
    for iw_section, iw_field, iw_keys in [
        ("infant_assessment", "weight_kg", [
            "infant_assessment.weight_kg", "infant_assessment.weight",
        ]),
        ("infant", "birth_weight_kg", [
            "infant.birth_weight_kg", "infant.birth_weight", "infant.weight",
        ]),
        ("child", "weight_kg", [
            "child.weight_kg", "child.weight",
        ]),
        ("growth_assessment", "weight_kg", [
            "growth_assessment.weight_kg", "growth_assessment.weight",
        ]),
    ]:
        for iw_key in iw_keys:
            iw_val = flat.get(iw_key)
            if iw_val is not None:
                section = result.get(iw_section, {})
                if isinstance(section, dict) and (iw_field not in section or section.get(iw_field) is None):
                    try:
                        num = float(str(iw_val).replace("kg", "").replace("KG", "").strip())
                        section[iw_field] = num
                        result[iw_section] = section
                    except (ValueError, TypeError):
                        pass
                break

    # ── Vitals weight normalization: "55 kg" → 55.0 ──
    # Only use vitals-specific keys to avoid grabbing infant weight
    for wkey in ("vitals.weight", "vitals.weight_kg"):
        wval = flat.get(wkey)
        if wval is not None:
            try:
                num = float(str(wval).replace("kg", "").replace("KG", "").strip())
                if "weight_kg" not in vitals or vitals.get("weight_kg") is None:
                    vitals["weight_kg"] = num
            except (ValueError, TypeError):
                pass
            break

    # ── Hemoglobin normalization ──
    for hkey in ("hemoglobin", "hemoglobin_gm_percent", "hb", "lab_results.hemoglobin"):
        hval = flat.get(hkey)
        if hval is not None:
            try:
                num = float(str(hval).replace("g/dl", "").replace("gm", "").strip())
                if "hemoglobin_gm_percent" not in vitals or vitals.get("hemoglobin_gm_percent") is None:
                    vitals["hemoglobin_gm_percent"] = num
            except (ValueError, TypeError):
                pass
            break

    if vitals:
        result["vitals"] = vitals

    # ── Gestational weeks normalization ──
    pregnancy = result.get("pregnancy", {})
    if "gestational_weeks" not in pregnancy or pregnancy.get("gestational_weeks") is None:
        for gkey in ("gestational_weeks", "gestational_age", "pregnancy.gestational_age",
                      "pregnancy.gestational_weeks", "gestation_weeks"):
            gval = flat.get(gkey)
            if gval is not None:
                try:
                    num = int(re.search(r'(\d+)', str(gval)).group(1))
                    pregnancy["gestational_weeks"] = num
                except (ValueError, TypeError, AttributeError):
                    pass
                break
    if pregnancy:
        result["pregnancy"] = pregnancy

    # ── Child age normalization ──
    for akey in ("age_months", "child.age_months", "age"):
        aval = flat.get(akey)
        if aval is not None:
            child = result.get("child", {})
            if isinstance(child, dict) and ("age_months" not in child or child.get("age_months") is None):
                try:
                    num = int(re.search(r'(\d+)', str(aval)).group(1))
                    child["age_months"] = num
                    result["child"] = child
                except (ValueError, TypeError, AttributeError):
                    pass
            break

    return result


def _run_inference_unsloth(system_prompt, user_prompt):
    """Run inference via Unsloth/transformers — slower but works without Ollama."""
    import torch
    model, tokenizer = load_model()

    messages = [
        {"role": "system", "content": system_prompt},
        {"role": "user", "content": user_prompt},
    ]
    text = tokenizer.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)
    inputs = tokenizer(text=[text], return_tensors="pt").to("cuda")

    t0 = time.time()
    with torch.no_grad():
        output_ids = model.generate(**inputs, max_new_tokens=768, do_sample=False)
    elapsed = time.time() - t0

    response = tokenizer.decode(output_ids[0][inputs["input_ids"].shape[-1]:], skip_special_tokens=True)
    parsed = _parse_json_response(response)
    return {"raw": response, "parsed": parsed, "time_s": elapsed}


def _parse_json_response(response):
    """Parse JSON from model response, handling markdown fences and quirks."""
    print(f"[DEBUG] raw response repr (first 80): {repr(response[:80])}")

    # Strip markdown fences — handle variations: ```json, ``` json, whitespace, BOM
    clean = response.strip().lstrip('\ufeff')
    clean = re.sub(r'^`{3,}\s*(?:json)?\s*[\r\n]*', '', clean, flags=re.IGNORECASE)
    clean = re.sub(r'[\r\n]*`{3,}\s*$', '', clean)
    clean = clean.strip()

    # Fix common model quirks
    if clean and clean[0] == '"' and not clean.startswith('{"') and not clean.startswith('["'):
        clean = "{" + clean
    if clean and clean[0] not in ('{', '['):
        first_brace = min(
            (clean.find("{") if clean.find("{") >= 0 else len(clean)),
            (clean.find("[") if clean.find("[") >= 0 else len(clean)),
        )
        if first_brace < len(clean):
            print(f"[DEBUG] skipped leading junk: {repr(clean[:first_brace])}")
            clean = clean[first_brace:]
    clean = re.sub(r'"{2,}([^"]+)"{2,}', r'"\1"', clean)
    clean = re.sub(r'(?<=: )"{2,}', '"', clean)
    clean = re.sub(r'"{2,}(?=\s*[,\}\]])', '"', clean)
    clean = re.sub(r',\s*([}\]])', r'\1', clean)

    print(f"[DEBUG] cleaned JSON (first 120): {repr(clean[:120])}")

    try:
        return json.loads(clean)
    except json.JSONDecodeError as e:
        print(f"[DEBUG] JSON parse failed: {e}")
        for end_pos in range(len(clean), max(0, len(clean) - 200), -1):
            if clean[end_pos - 1] in ('}', ']'):
                try:
                    parsed = json.loads(clean[:end_pos])
                    print(f"[DEBUG] recovered JSON by truncating at pos {end_pos}")
                    return parsed
                except json.JSONDecodeError:
                    continue

    print(f"[DEBUG] FULL raw response ({len(response)} chars):\n{response}\n---END---")
    return None


# ============================================================
# EXTRACTION PIPELINE
# ============================================================
def detect_visit_type(transcript):
    """Heuristic visit type detection from transcript content."""
    t = transcript.lower()
    # Delivery — check first, most specific keywords
    if any(kw in t for kw in ["डिलीवरी हो गई", "डिलीवरी हुई", "delivery हुई",
                               "डिलीवरी कब हुई", "delivery कब",
                               "जन्म हुआ", "पैदा हुआ", "प्रसव हुआ",
                               "लड़का हुआ", "लड़की हुई", "लड़की हुआ",
                               "घर पर ही हो गया", "घर पर हुई", "घर पर हुआ",
                               "ऑपरेशन से हुई", "caesarean", "सिजेरियन",
                               "जन्म का वजन", "birth weight", "birth_weight",
                               "जन्म के समय", "normal delivery", "दाई ने"]):
        return "delivery"
    # ANC — check before PNC/child (broad keywords like टीका overlap)
    if any(kw in t for kw in ["गर्भ", "प्रेग्नेंसी", "pregnancy", "anc", "पेट में बच्चा",
                               "गर्भवती", "हफ्ते की", "हफ्ते हो", "महीने की",
                               "lmp", "edd", "bp चेक", "hb ", "ifa", "tt का टीका",
                               "बच्चे की हलचल", "fetal", "डिलीवरी कहाँ", "डिलीवरी के लिए",
                               "जन्म के लिए तैयारी", "birth preparedness"]):
        return "anc_visit"
    # PNC — postpartum mother/newborn care
    if any(kw in t for kw in ["नवजात", "newborn", "दूध पीना", "दूध नहीं पीता", "दूध पीता",
                               "दूध पी रहा", "दूध नहीं पी", "दूध पिला",
                               "नाभि", "cord", "नाल", "स्तनपान",
                               "breastfeed", "imnci", "hbnc", "डिलीवरी के बाद",
                               "डिलीवरी को", "delivery को", "pnc",
                               "खून बहना", "खून आ रहा", "pad ", "पैड "]):
        return "pnc_visit"
    # Child health — older infants/children
    # Note: dropped "बच्चे को" — fires falsely on ANC danger-talk like
    # "तुम्हारा और बच्चे को खतरा" (preeclampsia warning to mother).
    # "child" also dropped — too generic, can appear in delivery/PNC counseling.
    if any(kw in t for kw in ["बच्चा कैसा", "बच्चा कैसी", "बच्चे का वजन", "बच्ची का वजन",
                               "टीका लग", "vaccine", "deworming", "vitamin a", "hbyc",
                               "महीने का", "महीने है", "दस्त", "diarrhea",
                               "खाता है", "खेलता है", "आँखें धँसी",
                               "सुस्त है", "सुस्त हो", "बहुत सुस्त"]):
        return "child_health"
    return "anc_visit"


def build_trimmed_danger_schema():
    """Danger sign schema without checklists — much smaller output."""
    return {
        "type": "object",
        "properties": {
            "visit_type": {
                "type": "string",
                "enum": ["antenatal", "postnatal_mother", "newborn", "child_under5"],
            },
            "danger_signs": {
                "type": "array",
                "description": "Detected danger signs. Empty array [] if none found.",
                "items": {
                    "type": "object",
                    "properties": {
                        "sign": {"type": "string"},
                        "category": {"type": "string", "enum": ["immediate_referral", "urgent_care", "monitor_closely"]},
                        "clinical_value": {"type": ["string", "null"]},
                        "utterance_evidence": {"type": "string", "description": "REQUIRED: exact verbatim quote"},
                    },
                    "required": ["sign", "category", "utterance_evidence"],
                },
            },
            "referral_decision": {
                "type": "object",
                "properties": {
                    "decision": {"type": "string", "enum": ["refer_immediately", "refer_within_24h", "continue_monitoring", "routine_followup"]},
                    "reason": {"type": "string"},
                },
                "required": ["decision", "reason"],
            },
        },
        "required": ["visit_type", "danger_signs", "referral_decision"],
    }


# Maternal danger sign names that map to checklist fields
MATERNAL_CHECKLIST_SIGNS = {
    "severe_vaginal_bleeding": ["vaginal bleeding", "severe bleeding", "रक्तस्राव", "खून"],
    "convulsions": ["convulsion", "seizure", "दौरा", "अकड़न"],
    # preeclampsia is the diagnostic name the LLM may emit instead of the symptom triad —
    # treat its presence as an explicit detection of severe headache + blurred vision
    "severe_headache_blurred_vision": [
        "headache", "blurred vision", "सिरदर्द", "धुंधला",
        "preeclampsia", "pre-eclampsia", "प्रीक्लिम्सिया", "प्री-एक्लेम्पसिया",
    ],
    "high_fever": ["high fever", "fever", "बुखार", "तेज़ बुखार"],
    "severe_abdominal_pain": ["abdominal pain", "पेट दर्द", "पेट में दर्द"],
    "fast_difficult_breathing": ["breathing", "साँस", "सांस"],
    # "सूज" matches the verb-stem (पैर सूज रहे हैं) which "सूजन" does not
    "swelling_face_hands": ["swelling", "edema", "सूजन", "सूज"],
    "reduced_fetal_movement": ["fetal movement", "reduced movement", "हलचल कम", "हिलता नहीं"],
    "water_break_prom": ["water break", "पानी टूट", "झिल्ली"],
    "foul_vaginal_discharge": ["discharge", "बदबूदार", "स्राव"],
}

NEWBORN_CHECKLIST_SIGNS = {
    "not_feeding_well": ["not feeding", "feeding", "दूध नहीं", "दूध पीना"],
    "convulsions": ["convulsion", "seizure", "दौरा"],
    "fast_breathing_gte60": ["fast breathing", "breathing", "साँस तेज़"],
    "severe_chest_indrawing": ["chest indrawing", "छाती धँसना"],
    "high_temperature": ["high temperature", "fever", "बुखार", "तापमान"],
    "low_temperature": ["low temperature", "ठंडा", "हाइपोथर्मिया"],
    "no_movement": ["no movement", "सुस्त", "हिलता नहीं"],
    "jaundice": ["jaundice", "पीलिया"],
    "umbilicus_red_pus": ["umbilicus", "नाभि", "cord"],
}


def derive_checklists(danger_signs, visit_type):
    """Derive maternal/newborn checklists from the danger_signs array."""
    maternal_ck = {k: "not_assessed" for k in MATERNAL_CHECKLIST_SIGNS}
    newborn_ck = {k: "not_assessed" for k in NEWBORN_CHECKLIST_SIGNS}

    if not danger_signs:
        return maternal_ck, newborn_ck

    # Check each detected sign against checklist keywords
    detected_signs_text = " ".join(
        f"{s.get('sign', '')} {s.get('utterance_evidence', '')}".lower()
        for s in danger_signs
    )

    for field, keywords in MATERNAL_CHECKLIST_SIGNS.items():
        if any(kw.lower() in detected_signs_text for kw in keywords):
            maternal_ck[field] = "detected"
        else:
            maternal_ck[field] = "not_detected"

    for field, keywords in NEWBORN_CHECKLIST_SIGNS.items():
        if any(kw.lower() in detected_signs_text for kw in keywords):
            newborn_ck[field] = "detected"
        else:
            newborn_ck[field] = "not_detected"

    return maternal_ck, newborn_ck


def validate_form_output(parsed, transcript):
    """Post-extraction validation: strip hallucinated fields, apply range checks.

    Common hallucination patterns on audio transcripts:
      - patient.name = "दीदी" / "बहन" / "Patient" (generic address, not a name)
      - patient.age = 30 (model's default guess)
      - lab_results.blood_group / hiv_status invented when not discussed
    """
    if not isinstance(parsed, dict):
        return parsed

    t_lower = transcript.lower() if transcript else ""

    # -- Name hallucination: generic Hindi address terms --
    FAKE_NAMES = {"दीदी", "बहन", "बहनजी", "patient", "दी दी", "didi", "bahen"}
    patient = parsed.get("patient") or {}
    name = patient.get("name") or patient.get("patient_name")
    if name and name.strip().lower() in FAKE_NAMES:
        if "patient" in parsed and isinstance(parsed["patient"], dict):
            for key in ("name", "patient_name"):
                if key in parsed["patient"]:
                    parsed["patient"][key] = None
                    print(f"[VALIDATE] Stripped hallucinated name: {name}")

    # -- Age hallucination: exactly 30 when not mentioned --
    age = patient.get("age") or patient.get("patient_age")
    if age == 30:
        # Check if "30" or "तीस" actually appears in transcript
        if "30" not in transcript and "तीस" not in transcript:
            if "patient" in parsed and isinstance(parsed["patient"], dict):
                for key in ("age", "patient_age"):
                    if key in parsed["patient"]:
                        parsed["patient"][key] = None
                        print(f"[VALIDATE] Stripped hallucinated age: 30")

    # -- Lab results hallucination: blood_group, HIV when not discussed --
    lab = parsed.get("lab_results") or {}
    BLOOD_GROUPS = {"a+", "a-", "b+", "b-", "ab+", "ab-", "o+", "o-"}
    bg = lab.get("blood_group")
    if bg and str(bg).strip().lower() in BLOOD_GROUPS:
        bg_mentioned = any(kw in t_lower for kw in ["blood group", "ब्लड ग्रुप", "खून का ग्रुप", "रक्त समूह"])
        if not bg_mentioned:
            parsed.setdefault("lab_results", {})["blood_group"] = None
            print(f"[VALIDATE] Stripped hallucinated blood_group: {bg}")

    hiv = lab.get("hiv_status") or lab.get("hiv")
    if hiv and str(hiv).strip().lower() in ("negative", "positive", "नेगेटिव", "पॉजिटिव"):
        hiv_mentioned = any(kw in t_lower for kw in ["hiv", "एचआईवी", "एड्स"])
        if not hiv_mentioned:
            for key in ("hiv_status", "hiv"):
                if key in parsed.get("lab_results", {}):
                    parsed["lab_results"][key] = None
                    print(f"[VALIDATE] Stripped hallucinated HIV: {hiv}")

    # -- Range checks on vital signs --
    RANGES = {
        "bp_systolic": (60, 250), "bp_diastolic": (30, 150),
        "weight_kg": (1, 200), "hemoglobin_gm_percent": (3, 20),
        "gestational_weeks": (1, 45), "temperature_f": (90, 110),
    }
    for section in [parsed, parsed.get("vitals", {}), parsed.get("pregnancy", {}),
                    parsed.get("anc_details", {}), parsed.get("newborn", {})]:
        if not isinstance(section, dict):
            continue
        for field, (lo, hi) in RANGES.items():
            val = section.get(field)
            if val is not None:
                try:
                    num = float(val)
                    if num < lo or num > hi:
                        section[field] = None
                        print(f"[VALIDATE] Out-of-range {field}={val} (valid: {lo}-{hi})")
                except (ValueError, TypeError):
                    pass

    return parsed


def extract_form(transcript, visit_type):
    """Extract structured form data from transcript."""
    schema = SCHEMAS.get(VISIT_TYPE_MAP.get(visit_type, "anc_visit"), SCHEMAS["anc_visit"])
    user_prompt = (
        f"Extract structured data from this ASHA home visit conversation:\n\n"
        f"{transcript}\n\n"
        f"Output JSON schema:\n{json.dumps(schema, ensure_ascii=False)}"
    )
    result = run_inference(FORM_SYSTEM_PROMPT, user_prompt)
    if result.get("parsed") and isinstance(result["parsed"], dict):
        result["parsed"] = validate_form_output(result["parsed"], transcript)
    return result


def extract_danger_signs(transcript, visit_type):
    """Extract danger signs using trimmed schema (no checklists) + post-validation."""
    schema = build_trimmed_danger_schema()
    user_prompt = (
        f"Analyze this ASHA home visit conversation for danger signs.\n\n"
        f"Visit type: {visit_type}\n\n"
        f"{transcript}\n\n"
        f"Output JSON schema:\n{json.dumps(schema, ensure_ascii=False)}"
    )
    result = run_inference(DANGER_SYSTEM_PROMPT, user_prompt)

    # Post-validation: drop danger signs whose evidence isn't in the transcript
    # or whose evidence is a generic ASHA phrase (not actual symptom description)
    GENERIC_PHRASES = [
        "कोई तकलीफ़ हो तो फ़ोन कर दीजिए",
        "कोई तकलीफ हो तो फोन कर दीजिए",
        "कोई समस्या हो तो तुरंत बताइए",
        "कोई समस्या हो तो फोन करें",
        "कोई दिक्कत हो तो",
        "अगली बार आऊँगी",
        "अगली विज़िट",
        "ठीक है दीदी, धन्यवाद",
        "ठीक है दीदी",
    ]

    # Normal vital sign readings that should NOT be flagged as danger signs
    NORMAL_INDICATORS = [
        "110/70", "120/80", "110/80", "118/76", "108/72",  # normal BP
        "बिल्कुल ठीक", "सामान्य", "नॉर्मल", "अच्छा है", "ठीक है",
        "बिल्कुल सामान्य",
    ]

    if result["parsed"] and "danger_signs" in result["parsed"]:
        validated_signs = []
        norm_transcript = re.sub(r'\s+', ' ', transcript.strip())

        for sign in result["parsed"]["danger_signs"]:
            evidence = sign.get("utterance_evidence", "")
            if not evidence or len(evidence) < 10:
                print(f"[DEBUG] dropped sign '{sign.get('sign','')}': evidence too short ({len(evidence)} chars)")
                continue

            norm_evidence = re.sub(r'\s+', ' ', evidence.strip())

            # Check against generic phrase blocklist
            is_generic = any(phrase in norm_evidence for phrase in GENERIC_PHRASES)
            if is_generic:
                print(f"[DEBUG] dropped sign '{sign.get('sign','')}': evidence is generic ASHA phrase")
                continue

            # Check if evidence describes a normal reading, not a danger sign
            is_normal = any(indicator in norm_evidence for indicator in NORMAL_INDICATORS)
            if is_normal:
                print(f"[DEBUG] dropped sign '{sign.get('sign','')}': evidence contains normal vital indicator")
                continue

            found = False
            if norm_evidence in norm_transcript:
                found = True
            elif len(norm_evidence) >= 20:
                min_chunk = min(30, len(norm_evidence))
                for i in range(0, len(norm_evidence) - min_chunk + 1):
                    chunk = norm_evidence[i:i + min_chunk]
                    if chunk in norm_transcript:
                        found = True
                        break

            if found:
                validated_signs.append(sign)
            else:
                print(f"[DEBUG] dropped sign '{sign.get('sign','')}': evidence not found in transcript")
                print(f"[DEBUG]   evidence: {repr(norm_evidence[:80])}")

        # If all remaining signs cite the same evidence, it's likely generic — drop all
        if len(validated_signs) > 1:
            evidences = set(s.get("utterance_evidence", "").strip() for s in validated_signs)
            if len(evidences) == 1:
                print(f"[DEBUG] dropped all {len(validated_signs)} signs: all cite same evidence (likely generic)")
                validated_signs = []

        dropped = len(result["parsed"]["danger_signs"]) - len(validated_signs)
        if dropped:
            print(f"[DEBUG] post-validation dropped {dropped}/{dropped + len(validated_signs)} danger signs")
        result["parsed"]["danger_signs"] = validated_signs

        if not validated_signs:
            result["parsed"]["referral_decision"] = {
                "decision": "routine_followup",
                "reason": "No danger signs detected in conversation",
            }

    # Derive checklists programmatically (instead of model generating them)
    if result["parsed"]:
        signs = result["parsed"].get("danger_signs", [])
        maternal_ck, newborn_ck = derive_checklists(signs, visit_type)
        result["parsed"]["maternal_danger_signs_checklist"] = maternal_ck
        result["parsed"]["newborn_danger_signs_checklist"] = newborn_ck

    return result


def _validate_fc_danger_signs(danger_signs, transcript):
    """Post-validate danger signs from function calling — same logic as extract_danger_signs."""
    GENERIC_PHRASES = [
        "कोई तकलीफ़ हो तो फ़ोन कर दीजिए",
        "कोई तकलीफ हो तो फोन कर दीजिए",
        "कोई समस्या हो तो तुरंत बताइए",
        "कोई समस्या हो तो फोन करें",
        "कोई दिक्कत हो तो",
        "अगली बार आऊँगी",
        "अगली विज़िट",
        "ठीक है दीदी, धन्यवाद",
        "ठीक है दीदी",
    ]
    NORMAL_INDICATORS = [
        "110/70", "120/80", "110/80", "118/76", "108/72",
        "बिल्कुल ठीक", "सामान्य", "नॉर्मल", "अच्छा है", "ठीक है",
        "बिल्कुल सामान्य",
    ]

    validated = []
    norm_transcript = re.sub(r'\s+', ' ', transcript.strip())

    for sign in danger_signs:
        evidence = sign.get("utterance_evidence") or sign.get("evidence", "")
        if not evidence or len(evidence) < 10:
            print(f"[DEBUG] FC dropped sign '{sign.get('sign','')}': evidence too short")
            continue

        norm_evidence = re.sub(r'\s+', ' ', evidence.strip())

        if any(phrase in norm_evidence for phrase in GENERIC_PHRASES):
            print(f"[DEBUG] FC dropped sign '{sign.get('sign','')}': generic phrase")
            continue
        if any(indicator in norm_evidence for indicator in NORMAL_INDICATORS):
            print(f"[DEBUG] FC dropped sign '{sign.get('sign','')}': normal vital")
            continue

        # Check evidence exists in transcript
        found = False
        if norm_evidence in norm_transcript:
            found = True
        elif len(norm_evidence) >= 20:
            min_chunk = min(30, len(norm_evidence))
            for i in range(0, len(norm_evidence) - min_chunk + 1):
                if norm_evidence[i:i + min_chunk] in norm_transcript:
                    found = True
                    break

        if found:
            validated.append(sign)
        else:
            print(f"[DEBUG] FC dropped sign '{sign.get('sign','')}': evidence not in transcript")

    # Same-evidence dedup
    if len(validated) > 1:
        evidences = set((s.get("utterance_evidence") or s.get("evidence", "")).strip() for s in validated)
        if len(evidences) == 1:
            print(f"[DEBUG] FC dropped all {len(validated)} signs: same evidence")
            validated = []

    dropped = len(danger_signs) - len(validated)
    if dropped:
        print(f"[DEBUG] FC post-validation dropped {dropped}/{len(danger_signs)} danger signs")
    return validated


def apply_metadata(form, visit_type, metadata):
    """Merge ASHA-entered patient identifier metadata into the LLM-extracted form.

    Metadata keys are schema-agnostic (patient_name, patient_age, age_unit, patient_sex,
    asha_id, visit_date, patient_mobile). This function overrides whichever schema-specific
    fields make sense for the visit type — leaving other LLM output untouched.

    PNC and delivery schemas have no patient block, so the metadata is preserved only
    in the envelope returned alongside the form (see extract_all).
    """
    if not form or not isinstance(form, dict) or not metadata:
        return form
    name = metadata.get("patient_name") or None
    age = metadata.get("patient_age")
    age_unit = (metadata.get("age_unit") or "").lower()
    sex = (metadata.get("patient_sex") or "").lower() or None
    mobile = metadata.get("patient_mobile") or None

    if visit_type == "anc_visit":
        patient = form.setdefault("patient", {}) if isinstance(form.get("patient"), dict) else None
        if patient is not None:
            if name: patient["name"] = name
            if age is not None and age_unit in ("", "years"):
                patient["age"] = age
            if mobile: patient["mobile"] = mobile
    elif visit_type == "child_health":
        child = form.setdefault("child", {}) if isinstance(form.get("child"), dict) else None
        if child is not None:
            if name: child["name"] = name
            if age is not None:
                # Convert to months for child_health schema
                if age_unit == "years":
                    child["age_months"] = int(age) * 12
                elif age_unit in ("", "months"):
                    child["age_months"] = int(age)
            if sex in ("male", "female"):
                child["sex"] = sex
    # pnc_visit and delivery — no schema-level patient block; envelope-only.
    return form


def extract_all(transcript, visit_type, metadata=None):
    """Hybrid extraction: format="json" for form (precise), function calling for danger+referral.
    Falls back to two format="json" calls if function calling is off.

    Optional `metadata` dict (patient identifier fields entered by ASHA before recording)
    is merged into the form and returned in the envelope. See apply_metadata().
    """
    if not (USE_OLLAMA and USE_FUNCTION_CALLING):
        # Fallback: two separate json-mode calls
        form_result = extract_form(transcript, visit_type)
        danger_result = extract_danger_signs(transcript, visit_type)
        form_data = apply_metadata(form_result.get("parsed"), visit_type, metadata)
        return {
            "form": form_data,
            "danger": danger_result.get("parsed"),
            "metadata": metadata or None,
            "tool_calls": [],
            "timing": {
                "form_s": round(form_result.get("time_s", 0), 1),
                "danger_s": round(danger_result.get("time_s", 0), 1),
            },
        }

    # ── Step 1: Form extraction via format="json" (proven precision) ──
    t0 = time.time()
    form_result = extract_form(transcript, visit_type)
    form_time = time.time() - t0
    form_data = form_result.get("parsed")

    # ── Step 2: Danger signs + referral via function calling ──
    fc_result = _run_danger_fc(transcript, visit_type)

    # Post-process danger signs
    raw_signs = fc_result["danger_signs"]
    validated_signs = _validate_fc_danger_signs(raw_signs, transcript)

    # Build referral decision
    referral_raw = fc_result["referral"]
    if validated_signs:
        urgency_map = {
            "immediate": "refer_immediately",
            "within_24h": "refer_within_24h",
            "routine": "continue_monitoring",
        }
        if referral_raw:
            referral_decision = {
                "decision": urgency_map.get(referral_raw.get("urgency"), "continue_monitoring"),
                "reason": referral_raw.get("reason", ""),
                "evidence_utterances": [s.get("utterance_evidence") or s.get("evidence", "") for s in validated_signs],
                "recommended_facility": referral_raw.get("facility"),
            }
        else:
            referral_decision = {
                "decision": "continue_monitoring",
                "reason": "Danger signs detected but no explicit referral issued",
                "evidence_utterances": [s.get("utterance_evidence") or s.get("evidence", "") for s in validated_signs],
            }
    else:
        referral_decision = {
            "decision": "routine_followup",
            "reason": "No danger signs detected in conversation",
            "evidence_utterances": [],
        }

    # Normalize danger sign format to match existing schema
    normalized_signs = []
    for s in validated_signs:
        normalized_signs.append({
            "sign": s.get("sign", ""),
            "category": s.get("category", "monitor_closely"),
            "clinical_value": s.get("clinical_value"),
            "utterance_evidence": s.get("utterance_evidence") or s.get("evidence", ""),
        })

    # Derive checklists
    maternal_ck, newborn_ck = derive_checklists(normalized_signs, visit_type)

    danger_data = {
        "visit_type": visit_type,
        "danger_signs": normalized_signs,
        "referral_decision": referral_decision,
        "maternal_danger_signs_checklist": maternal_ck,
        "newborn_danger_signs_checklist": newborn_ck,
    }

    form_data = apply_metadata(form_data, visit_type, metadata)

    return {
        "form": form_data,
        "danger": danger_data,
        "metadata": metadata or None,
        "tool_calls": fc_result["tool_calls"],
        "timing": {
            "form_s": round(form_time, 1),
            "danger_s": round(fc_result["time_s"], 1),
        },
    }