Skip to content

10 – Chat with Data

📘 RIDE User Manual – Panel 10: 🧠 Chat with Data (OpenAI + Hugging Face)

🎯 Purpose of the Panel

The Chat with Data panel empowers users to explore, analyze, and visualize their dataset through natural language prompts. It integrates with:

  • OpenAI GPT Models (e.g., gpt-4o, gpt-4o-mini)
  • Hugging Face Transformers (e.g., Mixtral, Mistral, Gemma, Falcon)

Users can ask questions about: - Dataset structure - Summary statistics - Missing values - Trends, correlations - Visualizations (bar, line, scatter, heatmaps) - Python code generation for transformations or EDA

🧭 User Workflow

  1. Upload Dataset

    • Upload CSV or Excel file.
    • The dataset is automatically read and shown in a preview.

  2. Choose Model Provider

    • 🧠 OpenAI (requires OpenAI API key)
    • 🤗 Hugging Face (requires Hugging Face API token)

  3. Select a DataFrame

    • Choose from initial, encoded, imputed, or scaled versions.

  4. Ask Questions

    • Examples:
      • “Which column has the most missing values?”
      • “Plot a bar chart of product categories.”
      • “How is age correlated with income?”

  5. Receive Answers

    • Get responses in plain English.
    • Includes Python code blocks if applicable.
    • Auto-executed plots rendered inline.

  6. Track Execution

    • Track query response times and performance metrics.

🤖 How It Works

🔐 OpenAI Mode

  • Integrates with openai.ChatCompletion.
  • Adds security filters to reject suspicious code (e.g., eval(), subprocess).
  • Uses safe_execute_code() to interpret and safely execute Python code blocks.
  • Tracks execution time and logs user queries.
  • Caches formatting using st.cache_data.

🤗 Hugging Face Mode

  • Uses InferenceClient from huggingface_hub.
  • Formats prompts using Mixtral-style [INST] tokens.
  • Directly queries LLMs with the full DataFrame context.
  • Parses responses, extracts Python code blocks and executes them for visualizations.

🧠 AI Safety and Security

Safety Feature Description
Suspicious Pattern Detection Blocks unsafe queries (e.g., drop table, open())
Resource Limit Guarding Automatically samples large datasets
Timeout Mechanism Halts long-running code after 30 seconds
Restricted Libraries Only pandas, numpy, matplotlib, seaborn permitted