fix

app.py

@@ -103,7 +103,8 @@ def build_llm_prompt(system: str, context: list[str], user_question: str) -> str
             prompt += f"[INST] {content.strip()} [/INST]\n"
         elif role == "assistant":
             prompt += f"{content.strip()}\n"
-    return prompt    
+    return prompt
+    
 
 # ---------- 4. Gradio playground (same UI as before) --------------------------
 def store_doc(doc_text: str, user_id="demo"):
@@ -218,6 +219,18 @@ with gr.Blocks() as demo:
         outputs=answer_box
     )
 
+# ---------- 3. FastAPI layer --------------------------------------------------
+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