feat(rag): honest deflection gate with amber UI state and alternative-question chips

app/api/chat.py

@@ -1,33 +1,39 @@
 """POST /api/chat — a RAG chat turn streamed over SSE (PLAN §3/§4).
 
 Flow (LOCKED A7/A15): embed the question → pgvector cosine top-K chunks →
-distinct parent documents (full text, capped) → locked persona prompt
-(PLAN §6) → ``turbo`` streamed as ``delta`` events → final ``done`` event
-(``deflected``, ``sources``, ``suggestions``) + ``query_log`` row + the
-per-turn log line (PLAN §9). Mid-stream failures become a structured
-``error`` event; a pre-stream DB outage is a plain 503 JSON.
+the **honesty gate** (A8: best score < ``BOR_RELEVANCE_THRESHOLD`` ⇒
+deflection) → locked persona prompt (PLAN §6) → ``turbo`` streamed as
+``delta`` events → final ``done`` event (``deflected``, ``sources``,
+``suggestions``) + ``query_log`` row + the per-turn log line (PLAN §9).
+Mid-stream failures become a structured ``error`` event; a pre-stream DB
+outage is a plain 503 JSON.
 
-The honesty gate (LOW relevance → deflection) lands in phase 04; every
-turn in this phase is grounded (``deflected=false``).
+Honesty gate: a weak retrieval (score strictly below the threshold — or
+an empty KB) flips the turn to deflection mode: the LOW prompt carries
+weak-hit *titles only* (never document content) plus deterministic
+"Maybe try" chips, and the ``done`` event / ``query_log`` row record
+``deflected=true`` with the weak score.
 """
 from __future__ import annotations
 
 import json
 import logging
 import time
-from collections.abc import AsyncIterator
+from collections.abc import AsyncIterator, Sequence
+from dataclasses import dataclass
 from typing import Any
 
 from fastapi import APIRouter, Depends
 from fastapi.responses import JSONResponse, StreamingResponse
 from sqlalchemy.orm import Session
 
-from app.config import get_settings
+from app.config import Settings, get_settings
 from app.db import db_available, get_db
-from app.models import QueryLog
+from app.models import Document, QueryLog
 from app.rag.llm import EmbeddingError, LLMClient, LLMError
-from app.rag.prompts import build_high_prompt
-from app.rag.retriever import retrieve, select_documents
+from app.rag.prompts import build_deflect_prompt, build_high_prompt
+from app.rag.retriever import RetrievedChunk, retrieve, select_documents, weak_hit_titles
+from app.rag.suggestions import derive_suggestions
 from app.schemas import ChatDoneEvent, ChatRequest, SourceRef
 
 logger = logging.getLogger("app.chat")
@@ -52,6 +58,44 @@ def sse_event(payload: dict[str, Any]) -> str:
     return f"data: {json.dumps(payload, ensure_ascii=False)}\n\n"
 
 
+@dataclass
+class TurnPlan:
+    """What one chat turn sends to the LLM and reports on ``done``."""
+
+    top_score: float
+    deflected: bool
+    system_prompt: str
+    docs: list[Document]  # cited sources (weak hits when deflected)
+    suggestions: list[str]  # "Maybe try" chips (deflected turns only)
+
+
+def plan_turn(chunks: Sequence[RetrievedChunk], settings: Settings) -> TurnPlan:
+    """Apply the honesty gate (A8) and assemble prompt + context for a turn.
+
+    * ``top_score >= threshold`` → grounded: HIGH prompt with the full
+      top-N documents, no suggestions. A score exactly at the threshold
+      is an answer — the gate is strict (``score < threshold``).
+    * ``top_score < threshold`` (or no hits at all) → deflected: LOW
+      prompt (``DEFLECT_MODE``) with weak-hit titles only — never document
+      content — plus deterministic alternative-question chips derived
+      from those titles.
+    """
+    top_score = chunks[0].score if chunks else 0.0
+    if top_score >= settings.relevance_threshold:
+        docs = select_documents(
+            chunks, n=settings.top_n_docs, max_chars=settings.max_context_chars
+        )
+        return TurnPlan(top_score, False, build_high_prompt(docs), docs, [])
+    titles = weak_hit_titles(chunks)
+    return TurnPlan(
+        top_score,
+        True,
+        build_deflect_prompt(titles),
+        select_documents(chunks, n=settings.top_n_docs, max_chars=settings.max_context_chars),
+        derive_suggestions(titles, settings.suggestions),
+    )
+
+
 @router.post("/chat")
 async def chat(
     request: ChatRequest,
@@ -88,10 +132,12 @@ async def chat(
             return
         embed_ms = int((time.monotonic() - t0) * 1000)
 
-        # 2. Retrieve top-K chunks → top-N full parent documents.
+        # 2. Retrieve top-K chunks, then the honesty gate (A8) picks the
+        #    HIGH (grounded) or LOW (deflected) prompt + context.
+        settings = get_settings()
         try:
             chunks = retrieve(db, question_vec)
-            docs = select_documents(chunks)
+            plan = plan_turn(chunks, settings)
         except Exception:  # noqa: BLE001 — DB failure mid-turn
             logger.exception("chat: retrieval failed question=%r", request.message)
             yield sse_event(
@@ -101,15 +147,13 @@ async def chat(
                 }
             )
             return
-
-        top_score = chunks[0].score if chunks else 0.0
-        source_paths = [f"{d.source}/{d.path}" for d in docs]
+        source_paths = [f"{d.source}/{d.path}" for d in plan.docs]
         messages = [
-            {"role": "system", "content": build_high_prompt(docs)},
+            {"role": "system", "content": plan.system_prompt},
             {"role": "user", "content": request.message},
         ]
 
-        # 3. Stream the grounded answer.
+        # 3. Stream the answer (grounded, or an honest deflection).
         try:
             async for piece in llm.chat_stream(messages):
                 yield sse_event({"type": "delta", "text": piece})
@@ -126,9 +170,9 @@ async def chat(
             db.add(
                 QueryLog(
                     question=request.message,
-                    top_score=top_score,
+                    top_score=plan.top_score,
                     chunk_hits=len(chunks),
-                    deflected=False,
+                    deflected=plan.deflected,
                     sources=", ".join(source_paths),
                     latency_ms=total_ms,
                 )
@@ -142,19 +186,19 @@ async def chat(
             "sources=%r total_ms=%d",
             request.message,
             embed_ms,
-            top_score,
-            get_settings().relevance_threshold,
-            False,
+            plan.top_score,
+            settings.relevance_threshold,
+            plan.deflected,
             source_paths,
             total_ms,
         )
         yield sse_event(
             ChatDoneEvent(
-                deflected=False,
+                deflected=plan.deflected,
                 sources=[
-                    SourceRef(source=d.source, path=d.path, title=d.title) for d in docs
+                    SourceRef(source=d.source, path=d.path, title=d.title) for d in plan.docs
                 ],
-                suggestions=[],
+                suggestions=plan.suggestions,
             ).model_dump()
         )