A safety-first, self-healing agent that turns reliably turns natural-language data-analysis questions into code.
Rajath Rajesh
Gemini 2.0 Flash (via langchain_google_genai)
(easily swappable for GroqCloud, OpenAI, etc.)
| Layer | Status | Highlights |
|---|---|---|
| Meta-agent | ✓ | Converts NL question → pandas/Matplotlib/Seaborn code |
Static guard (guard.py) |
✓ | AST-based syntax & safety checks, blocks dangerous imports/calls, validates column names, now permits columns created in-snippet. |
Sandbox (run_in_repl) |
✓ | Python-AST REPL in a background thread, 2-second timeout, matplotlib switched to Agg to avoid macOS GUI crashes. |
| Self-healing loop | ✓ | Guard → Sandbox; on error sends a repair prompt (incl. error text) to the Meta-agent which rewrites the code ; retries ≤ 3. |
| Cross-checker | ✓ | Second LLM pass judges semantic correctness; can trigger up to 2 extra repair attempts. Afterwords, one final call to cross-verify |
CLI (main.py) |
✓ | python3 main.py --data <file> --query "<question>" works for .txt, .csv, .parquet, .feather. |
| Dataset helpers | ✓ | uci_dataset_prep (UCI power data from txt loader) + generic loader. |
| Evaluation scaffold | ✓ | 30 NL queries (UCI Household Energy and Kaggle Titanic Dataset) for quick benchmarking. |
| Docs | ✓ | README, notebook, prompt templates, guard comments, and usage examples. |
# 0. Clone the repository
git clone https://github.com/Cygnus101/Safeframe-AI.git
cd Safeframe-AI
# ── Option A · one-liner setup ─────────────────────────────────────────
# (recreates .venv, activates it, installs requirements)
. ./setup_env.sh # ← the leading dot means “source”
# ── Option B · manual setup (if you prefer) ────────────────────────────
# python3 -m venv .venv
# source .venv/bin/activate
# pip install --upgrade pip
# pip install -r requirements.txt
# 1. Add your Gemini (or other) API key
echo "GOOGLE_API_KEY=<your-key>" > .env # replace <your-key>
# 2. Place your dataset
mkdir -p dataset
# copy or download any CSV/Parquet there, e.g.:
# cp /path/to/household_power_consumption.txt dataset/
# 3. Run a sample query
python3 main.py --data dataset/household_power_consumption.txt \
--query "What was the average active power consumption in March 2007?"