Update api.py

api.py

@@ -294,4 +294,11 @@ async def async_generator_wrapper(prompt: str):
 @app.get("/generate")
 async def generate(request: Request):
     prompt = request.query_params.get("prompt", "안녕하세요")
-    return StreamingResponse(async_generator_wrapper(prompt), media_type="text/plain")
+    response_text = respond(prompt)
+    
+    async def stream_response():
+        for chunk in split_response(response_text):
+            yield chunk
+            await asyncio.sleep(0.1)
+
+    return StreamingResponse(stream_response(), media_type="text/plain")