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

apps/web/app/_components/ChatPanel.tsx

@@ -20,7 +20,7 @@ import { cn } from "@duoc-thu/ui";
 interface ChatPanelProps {
   sessionId: string;
   initialQuery?: string;
-  onCitationClick?: (citation: Citation, index: number) => void;
+  onCitationClick?: (citation: Citation, index: number, allCitations: Citation[]) => void;
   onCitationsLoaded?: (citations: Citation[]) => void;
   activeCitationIndex?: number | null;
   className?: string;
@@ -63,6 +63,7 @@ export function ChatPanel({
   const [error, setError] = useState<string | null>(null);
   const messagesEndRef = useRef<HTMLDivElement>(null);
   const abortControllerRef = useRef<AbortController | null>(null);
+  const initialQuerySentRef = useRef<string | undefined>(undefined);
 
   const scrollToBottom = () => {
     messagesEndRef.current?.scrollIntoView({ behavior: "smooth" });
@@ -132,7 +133,15 @@ export function ChatPanel({
   };
 
   useEffect(() => {
-    if (initialQuery) {
+    // Guard against firing twice for the same query: React 18 Strict Mode
+    // (dev only) runs this effect setup twice on mount, and with no guard
+    // that sent every quick-prompt click as two identical live requests
+    // (found live 2026-08-07: duplicate "Chỉ định & Tác dụng không mong
+    // muốn của Aspirin" turns in the trace). The ref persists across the
+    // Strict Mode replay, so the second invocation for the same
+    // `initialQuery` is a no-op; a genuinely new query still sends once.
+    if (initialQuery && initialQuerySentRef.current !== initialQuery) {
+      initialQuerySentRef.current = initialQuery;
       handleSendMessage(initialQuery);
     }
     // eslint-disable-next-line react-hooks/exhaustive-deps
@@ -291,15 +300,31 @@ export function ChatPanel({
         {messages.length === 0 ? (
           renderEmptyState()
         ) : (
-          messages.map((msg) => (
-            <ChatBubble
-              key={msg.id}
-              message={msg}
-              onCitationClick={(citation, idx) => onCitationClick?.(citation, idx)}
-              activeCitationIndex={activeCitationIndex}
-              onRetry={() => handleSendMessage(msg.content)}
-            />
-          ))
+          messages.map((msg, msgIdx) => {
+            // Retry must resend the ORIGINAL user question, not this
+            // bubble's own text — for an assistant bubble, `msg.content` is
+            // the answer/error text itself, so resending it fed the error
+            // message back in as if it were the next question (found live
+            // 2026-08-07: a trace row where the query text WAS literally
+            // "Dịch vụ đang gặp sự cố tạm thời..."). Walk back to the
+            // nearest preceding user turn instead.
+            const retryQuery =
+              msg.role === "assistant"
+                ? [...messages.slice(0, msgIdx)].reverse().find((m) => m.role === "user")?.content
+                : undefined;
+            return (
+              <ChatBubble
+                key={msg.id}
+                message={msg}
+                onCitationClick={(citation, idx, allCitations) =>
+                  onCitationClick?.(citation, idx, allCitations)
+                }
+                activeCitationIndex={activeCitationIndex}
+                onRetry={retryQuery ? () => handleSendMessage(retryQuery) : undefined}
+                onQuickReply={(text) => handleSendMessage(text)}
+              />
+            );
+          })
         )}
 
         {/* Loading Indicator */}