refactor: Revise system prompt and enhance chat handling in the Gradio application by clarifying rules and workflow, improving file upload handling, and updating the chat response pipeline for better structure and clarity.

app.py

@@ -4,7 +4,6 @@ Gradio + OpenAI MCP Connector — Clean, Fast, Streaming, With File Upload
 
 import os
 import shutil
-
 import gradio as gr
 from openai import OpenAI
 
@@ -13,27 +12,41 @@ from openai import OpenAI
 # ---------------------
 OPENAI_API_KEY = os.getenv("OPENAI_API_KEY")
 MCP_SERVER_URL = "https://mcp-1st-birthday-auto-deployer.hf.space/gradio_api/mcp/"
-MODEL_FAST = "gpt-5-mini"  # for tool resolution
-MODEL_STREAM = "gpt-5.1"  # for final streaming reply
+
+MODEL_FAST = "gpt-5-mini"   # for tool resolution / MCP calls
+MODEL_STREAM = "gpt-5.1"       # for final streaming reply
 
 client = OpenAI(api_key=OPENAI_API_KEY)
 
+# ---------------------
+# SYSTEM PROMPT
+# ---------------------
 SYSTEM_PROMPT = """
-You are a fast MLOps assistant with access to remote MCP tools.
-Use tools only when necessary.
-Keep reasoning effort LOW for speed.
-After tools run, summarize clearly and concisely.
+You are a fast MLOps automation assistant equipped with remote MCP tools
+for dataset analysis, model training, evaluation, and deployment.
+
+Rules:
+- Use MCP tools when they directly address the user's request.
+- Treat the uploaded CSV file URL as the source of truth. Never modify it.
+- Never hallucinate tool names, arguments, or fields.
+- Keep your internal reasoning hidden.
+- Keep responses short, direct, and practical.
+
+Workflow:
+1) Decide if a tool is needed for the request.
+2) If needed, call the correct MCP tool with the exact schema.
+3) After tools complete, give a concise summary in plain language.
+4) If no tool is needed, answer directly and briefly.
 """
 
 # ---------------------
-# NATIVE MCP CONNECTOR
+# NATIVE MCP CONNECTOR (HTTP STREAMING)
 # ---------------------
 TOOLS = [
     {
         "type": "mcp",
         "server_label": "deploy_tools",
-        "server_url": MCP_SERVER_URL,
-        # transport auto-detected; HF space supports HTTP
+        "server_url": MCP_SERVER_URL,  # HTTP streaming MCP server
     }
 ]
 
@@ -42,139 +55,54 @@ TOOLS = [
 # FILE UPLOAD HANDLER
 # ---------------------
 def handle_upload(file_obj, request: gr.Request):
+    """
+    - Persist uploaded file to a stable /tmp path
+    - Return a public URL that the MCP tools can use directly
+    """
     if file_obj is None:
         return None
 
-    # Ensure file is in a stable path
     local_path = file_obj.name
     stable_path = os.path.join("/tmp", os.path.basename(local_path))
+
     try:
         shutil.copy(local_path, stable_path)
         local_path = stable_path
     except Exception:
+        # If copy fails, still try original path
         pass
 
-    # Build public Gradio URL
     base = str(request.base_url).rstrip("/")
     return f"{base}/gradio_api/file={local_path}"
 
 
 # ---------------------
-# MAIN CHAT HANDLER
+# MAIN CHAT HANDLER (STREAMING)
 # ---------------------
 def chat_send_stream(user_msg, history, file_url):
-    # history always starts as list of dicts
+    """
+    2-phase pipeline:
+      PHASE 1: Non-streaming tool resolution using MODEL_FAST
+      PHASE 2: Streaming final answer using MODEL_STREAM
+    Gradio Chatbot expects: list[{"role": "...", "content": "..."}]
+    """
+
+    # Ensure history is list[dict(role, content)]
     if history is None:
         history = []
 
-    # Append the user's message
+    # Append the user message to the UI history first
     history.append({"role": "user", "content": user_msg})
 
-    # Build OpenAI message history
+    # ---- Build messages for OpenAI (sanitize away metadata etc.) ----
     messages = [{"role": "system", "content": SYSTEM_PROMPT}]
-    messages.extend(history)
+    for msg in history:
+        role = msg.get("role")
+        content = msg.get("content", "")
+        if role in ("user", "assistant"):
+            messages.append({"role": role, "content": content})
 
-    # Inject file context
-    final_user_msg = user_msg
+    # Inject file context into *last* user message for the model
     if file_url:
-        final_user_msg = f"[Uploaded CSV file: {file_url}]\n\n{user_msg}"
-
-    messages[-1] = {"role": "user", "content": final_user_msg}
-
-    # ----------------------------------
-    # PHASE 1 — TOOL RESOLUTION
-    # ----------------------------------
-    tool_phase = client.responses.create(
-        model=MODEL_FAST,
-        reasoning={"effort": "low"},
-        tools=TOOLS,
-        instructions=SYSTEM_PROMPT,
-        input=messages,
-    )
-
-    tool_feedback = []
-
-    if tool_phase.output:
-        for item in tool_phase.output:
-            if item.type == "tool_call":
-                tool_feedback.append(f"🛠️ Used tool `{item.name}`.")
-            elif item.type == "tool_result":
-                tool_feedback.append(str(item.content))
-
-        if not tool_feedback:
-            tool_feedback.append("No MCP tools needed.")
-
-    else:
-        tool_feedback.append("No MCP tools needed.")
-
-    # Add assistant message with tool feedback
-    history.append({"role": "assistant", "content": "\n".join(tool_feedback)})
-
-    yield history
-
-    # ----------------------------------
-    # PHASE 2 — STREAMING FINAL ANSWER
-    # ----------------------------------
-    final_msg = history[-1]["content"] + "\n\n"
-    history[-1]["content"] = final_msg
-
-    stream = client.responses.create(
-        model=MODEL_STREAM,
-        reasoning={"effort": "low"},
-        instructions=SYSTEM_PROMPT,
-        input=messages + [history[-1]],
-        stream=True,
-    )
-
-    for ev in stream:
-        if ev.type == "response.output_text.delta":
-            final_msg += ev.delta
-            history[-1]["content"] = final_msg
-            yield history
-
-        elif ev.type == "response.completed":
-            break
-
-    stream.close()
-
-
-# ---------------------
-# GRADIO UI
-# ---------------------
-with gr.Blocks(title="MCP + GPT-5 — Fast Streaming MLOps Agent") as demo:
-    gr.Markdown("""
-    # 🚀 AI-Driven MLOps Agent (MCP-Powered)
-    - Upload a CSV file  
-    - Tools resolve instantly  
-    - Final answer streams smoothly  
-    """)
-
-    file_state = gr.State()
-
-    uploader = gr.File(label="Upload CSV file", type="filepath", file_count="single")
-
-    uploader.change(handle_upload, inputs=[uploader], outputs=[file_state])
-
-    chatbot = gr.Chatbot(label="Chat")
-    msg = gr.Textbox(label="Message")
-    send = gr.Button("Send")
-
-    send.click(
-        chat_send_stream,
-        inputs=[msg, chatbot, file_state],
-        outputs=[chatbot],
-    ).then(lambda: "", outputs=[msg])
-
-    msg.submit(
-        chat_send_stream,
-        inputs=[msg, chatbot, file_state],
-        outputs=[chatbot],
-    ).then(lambda: "", outputs=[msg])
-
-
-if __name__ == "__main__":
-    demo.queue().launch(
-        allowed_paths=["/tmp"],
-        show_error=True,
-        quiet=True,
-    )
+        last_user = messages[-1]
+        if last_user["role"] == "user":