Send patient context once per session, not on every turn

apps/web/app/_components/ChatPanel.tsx

@@ -156,6 +156,15 @@ export function ChatPanel({
   // Fetched once; `getPatientProfile()` itself returns null on a 401, so an
   // anonymous visitor never even attempts an authenticated call more than once.
   const patientProfileRef = useRef<PatientProfile | null>(null);
+  // Which sessionId has already had the patient-context clause sent. Prepending
+  // on every single turn (the original design) adds "Bệnh nhân X tuổi, Y kg."
+  // noise to queries that have nothing to do with dosing/patient-specific
+  // context — measured live: it pushed a plain ADR-listing question
+  // (unrelated to age/weight) from a correct answerable answer into
+  // evidence_insufficient. `rag/understanding.py`'s multi-turn merge already
+  // carries a stated fact forward, so sending it once per conversation is
+  // enough — this only resets when `sessionId` itself changes.
+  const patientContextSentForSessionRef = useRef<string | null>(null);
 
   useEffect(() => {
     getPatientProfile()
@@ -202,16 +211,22 @@ export function ChatPanel({
     }, 1000);
 
     try {
+      patientContextSentForSessionRef.current = sessionId;
       const res = await fetch("/api/chat", {
         method: "POST",
         headers: { "Content-Type": "application/json" },
         body: JSON.stringify({
-          // Prepends saved patient context (age/weight/renal/hepatic/allergy)
+          // Prepends saved patient context on the first turn of this session
-          // to what's actually sent, never to what's shown in the chat
+          // only — never to what's shown in the chat bubble above. The LLM
-          // bubble above — the LLM understanding step already extracts these
+          // understanding step extracts these fields from free text and
-          // fields from free text every turn (see rag/understanding.py), so
+          // carries them across the conversation (rag/understanding.py), so
-          // this reuses that exact path instead of adding a second one.
+          // repeating them on every later turn only adds noise to queries
-          content: withPatientContext(userText, patientProfileRef.current),
+          // that aren't patient-specific (measured: it broke a plain ADR
+          // lookup). See patientContextSentForSessionRef above.
+          content:
+            patientContextSentForSessionRef.current === sessionId
+              ? userText
+              : withPatientContext(userText, patientProfileRef.current),
           conversationId: sessionId,
           responseMode,
         }),