Triage v5 — narrower logic & capture fields

live supersedes v4 backend: services/triage/src/triage/v5_narrow.py  Reference for the Layer-1 capture + routing decision in v5.

← v4 architecture MSK pathway viewer Coverage matrix

TL;DR

The v5 narrower does both Layer-1 capture (extract 17 structured clinical fields from free text) and routing (commit to a pw_code) in a single LLM call per turn. It never rejects a patient and never hits a hard turn cap — information completeness drives commit, and frustration is the only forced exit. Tone formatting wraps every reply with conversation-aware Olly voice (no repeated openers, no em dashes, empathy preserved on frustrated turns).

The 17 capture fields

Every field is optional. If the patient didn't mention it, the LLM is required to leave it null rather than guess — except for a small set of "explicit implications" (e.g. "I twisted my ankle" → onset: sudden because twist is by definition instantaneous).

The decision flow

Every user turn runs narrow_step(). The LLM ranks pathway candidates AND extracts the 17 fields in one call. Then five branches decide what to do:

Patient archetypes (what triggers what)

Tone formatter (Ver-Stack 2)

Every narrower response is post-processed by enforce_tone() with the canonical Olly voice spec. The formatter receives:

• The raw assistant message produced by the narrower
• The patient's last user message
• The last 6 prior assistant turns — so it can enforce "vary the acknowledgement" / "use 'I'm sorry to hear' at most once per conversation"
• preserve_empathy flag — set when the narrower's response is deliberately empathic (frustrated or escalate_gp). Tells the formatter not to strip the acknowledgement line ("I know this can feel like a lot") and rewrite it as a generic informational sentence.

Output: minimally-edited text + was_rewritten: true/false. Original returned untouched if the LLM call fails (fail-open).

Audit trail

Every narrower turn writes a row to triage.turns with the full capture under tool_args.v5_capture. Same payload is also stamped on the HTTP response as X-Triage-V5-Capture-B64 (base64-encoded JSON) so the chat client's debug-modal feature can render it per turn.

{
  "turn": 2,
  "v5_action": "pathway",
  "v5_pw_code": "olly_msk_037",
  "v5_confidence": 0.95,
  "v5_capture": {
    "extracted_fields": {
      "presenting_complaint": "I twisted my ankle.",
      "anatomical_region": "ankle",
      "onset": "sudden",
      "duration": "yesterday",
      "trigger": "playing football",
      "pain_score": 8,
      "pain_character": null,
      ...
    },
    "domain": {"classification": "MSK", "confidence": 1.0},
    "red_flag_mentions": ["trauma"],
    "safeguarding_flags": [],
    "confidence_overall": 0.98,
    "context_completeness": 0.33,
    "commit_path": "confidence_fast"
  }
}

{
  "turn": 3,
  "v5_action": "ask",
  "v5_capture": {
    "extracted_fields": { ..., "mood_signal": "frustrated", ... },
    "context_completeness": 0.20,
    "gated_pw_code": "olly_back_mvp",
    "gated_confidence": 0.85,
    "frustration_count": 1,
    "empathic_override": true
  }
}

Frontend debug modal

The chat client (/root/triage-rn/app/demo/chat/index.tsx) shows an info chip below every bot bubble. Tap it → modal renders the per-turn triage_meta JSON. Once the narrower commits a pathway, the chip swaps from ⓘ to PW <code> as a visible commit marker.

The BFF (app/api/chat+api.ts) base64-decodes the X-Triage-V5-Capture-B64 response header and forwards as triage_meta in the ?json=true body.

Valkey state — per chat_id

Hard rules

• Narrower never rejects a patient. No out_of_scope action. Greetings, vague messages, off-topic text all enter the fact-finding loop. False rejects are worse than redundant clarification turns.
• No hard turn cap. Information drives commit (completeness floor or confidence fast-path). Frustration counter is the only forced exit.
• Don't guess fields. If the patient didn't mention it, the LLM leaves it null. The only exceptions are the explicit-implication patterns (twist → sudden, "slept like a baby" → sleep_impact false, etc.) documented in the extraction prompt.
• Never name a condition in differential options. No sciatica, stenosis, arthritis etc. Use symptom / region phrasing only.
• Never bulldoze a frustrated patient into RF screening. Even at turn cap, frustration triggers empathic ask, not pathway commit.
• Tone formatter runs on every reply. Including ask, escape, escalate_gp. No raw LLM output reaches the patient.

Tuning knobs (env-configurable)

Try it

CHAT_ID=$(uuidgen)
KEY="olly_triage_grant_35_..."   # fixture bearer

# Turn 1 — thin opener, expect ask
curl -sS -X POST http://127.0.0.1:4008/v5/chat/completions \
  -H "Authorization: Bearer $KEY" -H "Content-Type: application/json" \
  -d "{\"model\":\"olly-v5\",\"chat_id\":\"$CHAT_ID\",
       \"messages\":[{\"role\":\"user\",\"content\":\"I twisted my ankle\"}]}"

# Turn 2 — rich detail, expect commit_path: confidence_fast
curl -sS ... \
  -d "{\"model\":\"olly-v5\",\"chat_id\":\"$CHAT_ID\",
       \"messages\":[
         {\"role\":\"user\",\"content\":\"I twisted my ankle\"},
         {\"role\":\"assistant\",\"content\":\"...\"},
         {\"role\":\"user\",\"content\":\"yesterday playing football, pain is 8/10\"}
       ]}"

Files

← v4 architecture v2 (legacy) MSK pathway viewer