fastapi

app.py

@@ -271,8 +271,20 @@ with gr.Blocks() as demo:
     )
 
 # ---------- 3. FastAPI layer --------------------------------------------------
-@gr.on_api()
-def api_answer(system_prompt, context, question, user_id, history_flag, temp, top_p, top_k):
-    return answer(system_prompt, context, question, user_id, history_flag, temp, top_p, top_k)
-
-demo.launch(show_api=True, share=True)
+class IngestReq(BaseModel):
+    user_id:str
+    docs:list[str]
+
+class QueryReq(BaseModel):
+    user_id:str
+    question:str
+
+api = FastAPI()
+api = gr.mount_gradio_app(api, demo, path="/")
+
+# ---------- 5. run both (FastAPI + Gradio) -----------------------------------
+if __name__ == "__main__":
+    # launch Gradio on a background thread
+    demo.queue().launch(share=False, prevent_thread_lock=True)
+    # then start FastAPI (uvicorn blocks main thread)
+    uvicorn.run(api, host="0.0.0.0", port=8000)