Update app.py

app.py

@@ -30,7 +30,9 @@ class TTSRequest(BaseModel):
     text: str
 
 
-UPLOAD_DIR = "uploads"
+# ⚠️ HF requires persistent storage under /data
+UPLOAD_DIR = "/data/uploads"
+os.makedirs(UPLOAD_DIR, exist_ok=True)
 
 # =====================================================
 # HEALTH CHECK
@@ -47,17 +49,15 @@ def health():
 
 @app.post("/upload")
 async def upload_file(file: UploadFile = File(...)):
-    os.makedirs(UPLOAD_DIR, exist_ok=True)
-
     file_path = os.path.join(UPLOAD_DIR, file.filename)
 
     with open(file_path, "wb") as f:
         f.write(await file.read())
 
-    # Update retriever
+    # Update vector store
     update_retriever(file_path)
 
-    # 🔥 Rebuild LangGraph so RAG is active
+    # 🔥 Rebuild LangGraph so RAG becomes active
     rebuild_graph()
 
     return {
@@ -74,6 +74,8 @@ async def upload_file(file: UploadFile = File(...)):
 async def chat(message: str, session_id: str = "default"):
 
     async def event_generator():
+        buffer = ""
+
         async for chunk in app_graph.astream(
             {"messages": [HumanMessage(content=message)]},
             config={"configurable": {"thread_id": session_id}},
@@ -85,10 +87,15 @@ async def chat(message: str, session_id: str = "default"):
             msg = chunk[0] if isinstance(chunk, tuple) else chunk
 
             if hasattr(msg, "content") and msg.content:
-                cleaned = msg.content.strip()
-                if cleaned:
-                    yield f"data: {cleaned}\n\n"
+                buffer += msg.content
+
+                # Flush every ~150 characters (prevents broken tokens)
+                if len(buffer) > 150:
+                    yield f"data: {buffer.strip()}\n\n"
+                    buffer = ""
 
+        if buffer:
+            yield f"data: {buffer.strip()}\n\n"
 
     return StreamingResponse(
         event_generator(),