Update app_gradio.py

app_gradio.py

@@ -6,8 +6,8 @@ from threading import Lock
 import gradio as gr
 
 from megumin_agent.chat import ChatServices
-from megumin_agent.chat import chat_once
 from megumin_agent.chat import create_chat_services
+from megumin_agent.chat import stream_chat
 
 
 INITIAL_GREETING = "내 이름은 메구밍! 홍마족 제일의 마법사이자, 폭렬 마법을 펼치는 자!"
@@ -682,22 +682,38 @@ async def respond(
     session_id: str | None,
 ):
     if not message.strip():
-        return history, session_id, "", "", current_badge()
+        yield history, session_id, "", "", current_badge()
+        return
 
-    reply, active_session_id = await chat_once(
+    updated_history = list(history)
+    updated_history.append({"role": "user", "content": message})
+    updated_history.append({"role": "assistant", "content": ""})
+    yield updated_history, session_id, "", "답변 생성 중...", current_badge(message)
+
+    active_session_id = session_id
+    got_reply = False
+    async for partial_text, active_session_id in stream_chat(
         user_message=message,
         services=get_services(),
         session_id=session_id,
-    )
-    updated_history = list(history)
-    updated_history.append({"role": "user", "content": message})
-    updated_history.append(
-        {
+    ):
+        got_reply = True
+        updated_history[-1] = {"role": "assistant", "content": partial_text}
+        yield (
+            updated_history,
+            active_session_id,
+            "",
+            "답변 생성 중...",
+            current_badge(message, partial_text),
+        )
+
+    if not got_reply:
+        updated_history[-1] = {
             "role": "assistant",
-            "content": reply or "오늘은 마력의 흐름이 조금 불안정하군요. 잠시 후 다시 시도해 주시겠습니까?",
+            "content": "오늘은 마력의 흐름이 조금 불안정하군요. 잠시 후 다시 시도해 주시겠습니까?",
         }
-    )
-    return updated_history, active_session_id, "", "", current_badge(
+
+    yield updated_history, active_session_id, "", "", current_badge(
         message,
         updated_history[-1]["content"],
     )