📊 AutoAnalyst

An agent that writes and runs its own data analysis.
Ask a question in plain English → AutoAnalyst writes Python, runs it against your data, reads the result, draws charts, and explains the answer — one step at a time.

What is this?

AutoAnalyst is a local, open-source take on ChatGPT's Advanced Data Analysis. The difference from a normal "ask the LLM" tool is the agent loop: instead of guessing an answer in one shot, the model is given a single tool — run_python — and a live Python session with your DataFrame already loaded. It writes code, runs it, sees the real output, and decides what to do next, repeating until it can answer.

Because every number it reports comes from code that actually executed against your data, the answers are grounded — not hallucinated.

The agent loop

flowchart LR
    Q["question + data schema"] --> M{{"Groq · llama-3.3-70b"}}
    M -- "run_python(code)" --> S["🐍 sandbox<br/>df · pd · np · plt"]
    S -- "stdout · errors · charts" --> M
    M -- "no tool call" --> A["plain-English answer"]

• One tool, real execution. The model writes pandas/matplotlib; the sandbox runs it.
• Persistent session. Variables (and df) survive across steps, like a notebook.
• Self-correcting. If code raises, the error is fed back and the model fixes it.
• Charts come back as images; the model only hears "a chart was rendered."

▶ Live demo

Try it on the Hugging Face Space — pick a dataset, ask a question, and watch the agent think in real time (the UI streams each step over Server-Sent Events).

Quickstart

Install

pip install autoanalyst

Set a free Groq key (put it in a .env file or your env):

GROQ_API_KEY=gsk_...

Use the CLI

autoanalyst -q "What was the survival rate by passenger class?" --csv titanic.csv

You'll see each step it runs — the code, the output, chart markers — then the answer.

Run the web demo locally

pip install "autoanalyst[serve]"
uvicorn app.server:app --port 8000
# open http://localhost:8000

Locally you can upload your own CSV/Excel; the public Space ships curated samples only.

How it works

Sample datasets

Four bundled, recognizable datasets (built by scripts/make_samples.py): Titanic (survival), restaurant tips, Palmer penguins, and a generated e-commerce sales table — each with a few verified example questions.

Safety

LLM-written code is executed, so the sandbox applies a pragmatic guard: an AST denylist (no os/subprocess/socket/file or network access), a per-snippet timeout, and capped output. It is not a bulletproof jail — what actually contains the risk is that the public demo runs only curated datasets inside a non-root container, and file upload is disabled there. Run untrusted analyses locally at your own discretion.

Limitations

• Built on an open 70B model via Groq — tool-calling is good but occasionally needs a retry on a malformed call (handled with a step cap and a forced final answer).
• It answers from the data you give it; it won't fetch external context.
• Groq geo-restricts some regions; the hosted Space runs where Groq is reachable.

Tech

QLoRA-free and GPU-free by design: Groq API · pandas · numpy · matplotlib · FastAPI · Server-Sent Events. Tested (GPU- and network-free) in CI.

License

MIT — see LICENSE.