Fix ai-service Dockerfile: bake in drug_entities.json, override its path

apps/ai-service/routers/rag.py

@@ -35,6 +35,9 @@ class CitationResponse(BaseModel):
     bbox: tuple[float, float, float, float] | None = None
     source_crop: str | None = None
     attachment: str | None = None
+    # 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 = ""
 
 
 class RagQueryResponse(BaseModel):
@@ -44,6 +47,16 @@ class RagQueryResponse(BaseModel):
     answer: str | None
     resolved_drug_id: str | None
     citations: list[CitationResponse]
+    # Whether `answer` is an LLM paraphrase (verified by grounding +
+    # entailment) or a verbatim extractive quote of the retrieved source —
+    # the UI shows these differently so a clinician knows which they're
+    # reading.
+    generated: bool = False
+    # Short suggested replies for a `decision == "clarify"` turn (e.g.
+    # ["Người lớn", "Trẻ em"]) — only populated by the sufficiency-check
+    # clarify path today; other clarify sources (no_drug, dosing_calc's
+    # needs_clarify) leave this empty rather than fabricate options.
+    quick_replies: list[str] = []
 
 
 def _answer_service(request: Request) -> GroundedAnswerService:
@@ -91,6 +104,7 @@ def _map_citations(items) -> list[CitationResponse]:
             bbox=item.bbox,
             source_crop=item.source_crop,
             attachment=item.attachment,
+            evidence_text=item.evidence_text,
         )
         for item in items
     ]
@@ -128,6 +142,8 @@ def query_rag(
         answer = reply.clarification if reply.clarification is not None else reply.answer
         resolved_drug_id = ", ".join(reply.drugs) if reply.drugs else None
         citations = _map_citations(reply.citations)
+        generated = reply.generated
+        quick_replies = list(reply.quick_replies)
     else:
         # No generator configured (ANSWER_PROVIDER=disabled): there is no LLM
         # to understand a turn with, so this is retrieval-only, single-turn,
@@ -138,12 +154,16 @@ def query_rag(
             answer = grounded.clarification
             resolved_drug_id = grounded.result.resolved_drug_id
             citations = []
+            generated = False
+            quick_replies = list(grounded.quick_replies)
         else:
             decision = grounded.result.decision.value
             reason = grounded.result.reason
             answer = grounded.answer
             resolved_drug_id = grounded.result.resolved_drug_id
             citations = _map_citations(grounded.citations)
+            generated = grounded.generated
+            quick_replies = []
 
     # Trace persistence is fail-open (F-09): an already-computed, safe answer
     # must reach the caller even if Postgres is unreachable. `save()` opens a
@@ -175,4 +195,6 @@ def query_rag(
         answer=answer,
         resolved_drug_id=resolved_drug_id,
         citations=citations,
+        generated=generated,
+        quick_replies=quick_replies,
     )