DataGuard-AI

A local data-profiling prototype for CSV and Excel files. It combines pandas and ydata-profiling reports with Ollama-generated suggestions for data-cleaning code. The repository contains a Streamlit interface and a separate FastAPI backend; it does not include the previously documented index.html frontend.

Important safety limit

Applying a fix executes Python on the host, without a sandbox. The API's apply_fix, the Streamlit AI-fix handler, and the custom-code panel call exec. A copied DataFrame does not isolate that code from files, processes, or the network. AI suggestions are untrusted code, not validated repairs.

Use only disposable, non-sensitive data in an isolated development environment. Do not expose the application to untrusted users or use it as a production data-cleaning service. Profiling reports can contain source values; treat uploads, generated reports, and downloads accordingly.

Local setup

Create a fresh environment rather than reusing the committed .venv:

git clone https://github.com/Zinga18018/DataGuard-AI.git
cd DataGuard-AI
python -m venv .venv-local
source .venv-local/bin/activate
python -m pip install -r requirements.txt

On Windows, activate with .venv-local\Scripts\activate. The requirements are unpinned. Excel .xlsx support also needs openpyxl, which is not listed in the requirements file.

For AI suggestions, provision Ollama and its mistral model separately. The code sends requests to http://localhost:11434/api/generate; provisioning a model requires a download and sufficient local resources. Neither interface is a security boundary around model-generated code.

Choose an interface from the repository root:

# Streamlit interface
python -m streamlit run app.py --server.address 127.0.0.1

# Or the separate FastAPI backend
python -m uvicorn api:app --host 127.0.0.1 --port 8000

The API documentation is at http://127.0.0.1:8000/docs. The Streamlit interface handles uploads, profiling, suggestions, and exports itself; it is not a frontend for that API.

Source map and limits

• app.py: Streamlit upload, profiling, custom-code, and download paths.
• ai_integration.py: Ollama suggestions and the AI-fix execution handler.
• api.py: upload, profile, recommendation, and fix endpoints.
• requirements.txt: declared Python dependencies.

HTML report export is implemented. PDF export additionally imports pdfkit (not declared) and uses a hard-coded Windows wkhtmltopdf executable path; installing the tool on PATH alone does not make this portable.

There is no committed project test suite or benchmark establishing large-dataset performance, correctness of suggested fixes, or secure isolation. Existing sample files and profiling reports are not documented as a vetted synthetic fixture. No project license file is included at the repository root.