refactor: Revise system prompt in Gradio application to emphasize concise, actionable summaries and structured output formatting, enhancing user interaction and clarity during data-related requests.

app.py

@@ -3,9 +3,9 @@ Gradio + OpenAI Responses API + Remote MCP Server (HTTP)
 CSV-based MLOps Agent with streaming final answer & MCP tools
 """
 
+import json
 import os
 import shutil
-import json
 
 import gradio as gr
 from openai import OpenAI
@@ -37,24 +37,112 @@ MAIN_SYSTEM_PROMPT = """
 You are a helpful MLOps assistant with MCP tools for CSV analysis, training,
 evaluation, and deployment.
 
-For data-related requests (datasets, CSVs, models, training, evaluation, 
-deployment), call MCP tools to get comprehensive natural language results.
-The tools will return detailed explanations you can share directly.
+Your primary goal is to give the user a SHORT, ACTIONABLE summary of what matters.
+Do NOT paste long tool outputs by default.
+
+You have access to MCP tools for:
+- CSV analysis
+- Model training
+- Evaluation
+- Deployment
+
+Use them when the user asks for anything related to datasets, CSVs, models,
+training, evaluation, predictions, or deployment.
+
+────────────────────────────────────
+OUTPUT FORMAT (VERY IMPORTANT)
+────────────────────────────────────
+
+Always structure your final answer in this exact order:
+
+1) A short **Key Summary** section (this is what should be streamed first):
+
+   - Start with the heading: `## Key Summary`
+   - Then give **3–7 bullet points** that cover:
+     - What you did (e.g. data analysis, training, evaluation, deployment)
+     - The most important metrics or outcomes
+     - Any critical warnings / caveats
+     - Concrete next steps for the user
+
+   - Keep this section:
+     - Concise
+     - High-signal
+     - Free of long logs, full tables, or raw JSON
+
+   - Do NOT include tool request/response payloads, HTTP URLs, or internal
+     route details here.
+
+2) An OPTIONAL collapsible **Tools & Technical Details** section:
+
+   Only include this if:
+   - Tools were actually used **and**
+   - The user has asked for details / config / logs OR you think more context
+     is truly important for them.
+
+   Use an HTML `<details>` block so that it is collapsible in the Gradio chatbot:
+
+   <details>
+     <summary>Show tools & technical details</summary>
+
+     - **MCP server label**: `auto-deployer`
+     - **MCP server URL**:
+       `https://mcp-1st-birthday-auto-deployer.hf.space/gradio_api/mcp/`
+
+     - **Tools used**
+       - Name / type (e.g. CSV analysis, training, evaluation, deployment)
+       - A one-line description of what each tool did.
+
+     - **Key parameters**
+       - Briefly list important arguments (e.g. target column, task type,
+         training options) as a short bullet list.
+
+     - **Important logs or outputs (optional)**
+       - Include only short, relevant snippets or summaries.
+       - If you show structured data, wrap it in fenced code blocks, e.g.:
+
+       ```json
+       {
+         "metric": "accuracy",
+         "value": 0.8732
+       }
+       ```
+
+   </details>
+
+   Inside this `<details>` block you may:
+   - Show tool names
+   - Show parameters
+   - Show MCP routes / URLs
+   - Show short log snippets or small JSON dumps
+
+   But still avoid dumping extremely long raw outputs unless the user
+   explicitly requests them.
 
-For general chat (no data operations), respond helpfully and naturally.
+────────────────────────────────────
+BEHAVIOR
+────────────────────────────────────
 
-When using tools:
-- Use the CSV file URL exactly as provided
-- Do not invent tool parameters  
-- Share the complete results from MCP tools
-- Add brief context or suggestions if helpful
+- For data-related requests (datasets, CSVs, models, training, evaluation,
+  deployment):
+  - Call MCP tools as needed.
+  - Read their full output.
+  - Distill everything into the `## Key Summary` section.
+  - Optionally add the collapsible `<details>` block if you feel it is helpful
+    or the user asked for it.
 
-Keep responses clear, informative, and user-friendly.
+- For general, non-data chat:
+  - You can still use `## Key Summary` for clarity.
+  - You may omit the `<details>` block if no tools are used.
 
-Formatting rules:
-- Use Markdown for formatting
-- Use bullet points for lists
-- Wrap code, commands, and JSON in fenced code blocks
+────────────────────────────────────
+FORMATTING RULES
+────────────────────────────────────
+
+- Use Markdown for headings, lists, and emphasis.
+- Use bullet points for lists.
+- Wrap code, commands, and JSON in fenced code blocks.
+- Always output `## Key Summary` first so that the streaming response gives
+  the user the high-level picture before any technical details.
 """
 
 
@@ -86,21 +174,27 @@ def extract_output_text(response) -> str:
     Extract text from a non-streaming Responses API call while preserving formatting.
     """
     try:
-        if hasattr(response, 'output') and response.output and len(response.output) > 0:
+        if hasattr(response, "output") and response.output and len(response.output) > 0:
             first = response.output[0]
             if getattr(first, "content", None):
                 for content_item in first.content:
-                    if hasattr(content_item, 'type') and content_item.type == "output_text":
+                    if (
+                        hasattr(content_item, "type")
+                        and content_item.type == "output_text"
+                    ):
                         text = getattr(content_item, "text", None)
                         if text:
                             return text
-                    elif hasattr(content_item, 'type') and content_item.type == "output_json":
+                    elif (
+                        hasattr(content_item, "type")
+                        and content_item.type == "output_json"
+                    ):
                         # If there's JSON output, format it nicely
-                        json_data = getattr(content_item, 'json', None)
+                        json_data = getattr(content_item, "json", None)
                         if json_data:
                             return f"```json\n{json.dumps(json_data, indent=2)}\n```"
         # Fallback
-        return getattr(response, 'output_text', None) or str(response)
+        return getattr(response, "output_text", None) or str(response)
     except Exception as e:
         return f"Error extracting output: {e}"