Checkpoint frontend UI/UX overhaul and ingestion embed benchmark work

apps/ai-service/rag/conversational.py

@@ -343,10 +343,31 @@ class ConversationalLoopService:
             query, subject_scope, intent, drug_id=resolved.drug_id
         )
 
+        # The engine asked for a missing detail (age/weight/renal function/
+        # indication…) rather than dumping every dose band — surface it as a
+        # clarification, not a final answer.
+        if grounded is not None and grounded.clarification is not None:
+            self._metrics.increment(
+                metric_names.CLARIFY_ASKED, reason="needs_more_info"
+            )
+            self._persist(state, resolved, None)
+            return ConversationTurnResult(
+                None,
+                Clarification(
+                    reason="needs_more_info",
+                    question=grounded.clarification,
+                    options=(),
+                ),
+                None,
+                False,
+                None,
+                "needs_more_info",
+            )
+
         answer = grounded.answer if grounded else None
         inherited = resolved.drug_id if resolved.needs_carry_over_notice else None
         if answer is not None and inherited is not None:
-            answer = f"Về {inherited}: {answer}"
+            answer = f"Về {self._drug_name(inherited)}: {answer}"
         if grounded is not None:
             grounded = replace(grounded, answer=answer)
 
@@ -395,5 +416,9 @@ class ConversationalLoopService:
             and state.overflow()
         ):
             summary = self._summariser.fold(state.summary, state.overflow())
-            state = replace(state, summary=summary)
+            # Clear what was just folded — `replace()` keeps every field not
+            # named here, and `pending_overflow` accumulates across calls
+            # (see `ConversationState.append`), so leaving it would fold the
+            # same already-summarised turns again next time.
+            state = replace(state, summary=summary, pending_overflow=())
         self._store.save(state)