Add production condition retrieval smoke test

apps/ai-service/routers/rag.py

@@ -1,12 +1,12 @@
 from __future__ import annotations
 
 import uuid
-from typing import Annotated, Any, Protocol
+from typing import Annotated, Any, Literal, Protocol
 
 from fastapi import APIRouter, Depends, HTTPException, Request
 from pydantic import BaseModel, Field
 
-from rag.answer import GroundedAnswerService
+from rag.answer import DISCLAIMER, GroundedAnswerService
 from rag.metrics import DECISION, TRACE_WRITE_FAILED, Metrics, NullMetrics
 from rag.models import QueryIntent, SubjectScope
 from rag.policy import resolve_subject_scope
@@ -16,11 +16,14 @@ from rag.telemetry import (
     current_trace_id,
     stage,
 )
+from adapters.postgres import FeedbackTraceNotFound
 
 
 class TraceWriter(Protocol):
     def save(self, **fields: Any) -> str: ...
 
+    def save_feedback(self, **fields: Any) -> str: ...
+
 
 class RagQueryRequest(BaseModel):
     query: str = Field(min_length=1, max_length=4000)
@@ -44,6 +47,11 @@ class CitationResponse(BaseModel):
     # The exact retrieved chunk text this citation stands for — lets the UI
     # show precisely what was retrieved, not a client-side guess at it.
     evidence_text: str = ""
+    drug_id: str | None = None
+    drug_name: str | None = None
+    section_key: str | None = None
+    section_title: str | None = None
+    source_document: str = "Dược thư Quốc gia Việt Nam 2018"
 
 
 class AnswerClaimResponse(BaseModel):
@@ -65,6 +73,27 @@ class AnswerPlanResponse(BaseModel):
     needs_warning: bool
 
 
+class MedicationCandidateAssessmentResponse(BaseModel):
+    drug_id: str
+    drug_name: str
+    indication_supported: bool
+    status: str
+    indication_source_ids: list[str]
+    safety_source_ids: list[str]
+
+
+class RagFeedbackRequest(BaseModel):
+    trace_id: uuid.UUID
+    rating: Literal["helpful", "not_helpful"]
+    comment: str | None = Field(default=None, max_length=2000)
+    conversation_id: str | None = Field(default=None, max_length=128)
+
+
+class RagFeedbackResponse(BaseModel):
+    feedback_id: str
+    status: Literal["saved"] = "saved"
+
+
 class RagQueryResponse(BaseModel):
     trace_id: str
     correlation_id: str
@@ -87,6 +116,15 @@ class RagQueryResponse(BaseModel):
     blocks: list[AnswerBlockResponse] = []
     answer_mode: str = "concise"
     answer_plan: AnswerPlanResponse | None = None
+    candidate_assessments: list[MedicationCandidateAssessmentResponse] = []
+    # `docs/architecture.md` specifies the disclaimer at several layers. The
+    # web banner was the only one in place, so any other consumer received
+    # medical content with nothing attached. It is a fixed string owned by
+    # `rag/answer.py`, never produced by the model, and it is present on
+    # every decision — an abstain or a clarification is still a clinical
+    # response. Defaulted here as well so a response constructed in a test or
+    # a future code path cannot accidentally omit it.
+    disclaimer: str = DISCLAIMER
 
 
 def _answer_service(request: Request) -> GroundedAnswerService:
@@ -110,6 +148,26 @@ def _metrics(request: Request) -> Metrics:
 router = APIRouter(prefix="/v1/rag", tags=["rag"])
 
 
+@router.post("/feedback", response_model=RagFeedbackResponse)
+def save_feedback(
+    payload: RagFeedbackRequest,
+    traces: Annotated[TraceWriter, Depends(_trace_writer)],
+) -> RagFeedbackResponse:
+    comment = payload.comment.strip() if payload.comment else None
+    try:
+        feedback_id = traces.save_feedback(
+            trace_id=str(payload.trace_id),
+            rating=payload.rating,
+            comment=comment or None,
+            conversation_id=payload.conversation_id,
+        )
+    except FeedbackTraceNotFound as exc:
+        raise HTTPException(status_code=404, detail="trace_not_found") from exc
+    except Exception as exc:
+        raise HTTPException(status_code=503, detail="feedback_store_unavailable") from exc
+    return RagFeedbackResponse(feedback_id=feedback_id)
+
+
 class SuggestResponse(BaseModel):
     suggestions: list[str]
 
@@ -135,6 +193,11 @@ def _map_citations(items) -> list[CitationResponse]:
             source_crop=item.source_crop,
             attachment=item.attachment,
             evidence_text=item.evidence_text,
+            drug_id=item.drug_id,
+            drug_name=item.drug_name,
+            section_key=item.section_key,
+            section_title=item.section_title,
+            source_document=item.source_document,
         )
         for item in items
     ]
@@ -166,6 +229,26 @@ def _map_plan(item) -> AnswerPlanResponse | None:
     )
 
 
+def _map_candidate_assessments(items) -> list[MedicationCandidateAssessmentResponse]:
+    return [
+        MedicationCandidateAssessmentResponse(
+            drug_id=item.drug_id,
+            drug_name=item.drug_name,
+            indication_supported=item.indication_supported,
+            status=item.status.value,
+            indication_source_ids=[
+                evidence.matched_doc_id for evidence in item.indication_evidence
+            ],
+            safety_source_ids=[
+                evidence.matched_doc_id
+                for evidence in item.evidence
+                if evidence not in item.indication_evidence
+            ],
+        )
+        for item in items
+    ]
+
+
 @router.post("/query", response_model=RagQueryResponse)
 def query_rag(
     payload: RagQueryRequest,
@@ -203,6 +286,9 @@ def query_rag(
         blocks = _map_blocks(reply.blocks)
         answer_mode = reply.answer_mode
         answer_plan = _map_plan(reply.plan)
+        candidate_assessments = _map_candidate_assessments(
+            reply.candidate_assessments
+        )
     else:
         # No generator configured (ANSWER_PROVIDER=disabled): there is no LLM
         # to understand a turn with, so this is retrieval-only, single-turn,
@@ -219,6 +305,7 @@ def query_rag(
             blocks = []
             answer_mode = "concise"
             answer_plan = None
+            candidate_assessments = []
         else:
             decision = grounded.result.decision.value
             reason = grounded.result.reason
@@ -230,6 +317,7 @@ def query_rag(
             blocks = _map_blocks(grounded.blocks)
             answer_mode = grounded.answer_mode
             answer_plan = _map_plan(grounded.plan)
+            candidate_assessments = []
 
     # Trace persistence is fail-open (F-09): an already-computed, safe answer
     # must reach the caller even if Postgres is unreachable. `save()` opens a
@@ -284,5 +372,6 @@ def query_rag(
             blocks=blocks,
             answer_mode=answer_mode,
             answer_plan=answer_plan,
+            candidate_assessments=candidate_assessments,
         )
     return response