RAGtag is a pre-ingestion gate for RAG knowledge bases. It scores candidate
documents for embedding anomaly, instruction injection, and retrieval influence,
then returns an ADMIT, QUARANTINE, or REJECT verdict.
The repository includes the local FAISS/Ollama demo, a compatible-endpoint adapter, three poisoning signals, signed admission evidence, an asynchronous API, a Streamlit analyst dashboard, and a 60-document labelled evaluation set.
- Python 3.11
- Ollama for the offline demo model
python3.11 -m venv .venv && source .venv/bin/activate
pip install -r requirements.txt
ollama pull phi3:mini # or gemma2:2bragtag seed # build corpus index + cache clean probe answers
ragtag scan <file> # score one document, print verdict
ragtag --demo-mode scan <file> # pre-warm caches/model before the live scan
ragtag verify <file> <report> # offline integrity check
ragtag eval # metrics + data/labelled/eval_results.json
uvicorn ragtag.api:app --reload
streamlit run dashboard/app.py
pytest -qWhen the console script is not installed, use python -m ragtag.cli in place
of ragtag.
The API, dashboard, Ollama, and phi3:mini model start together:
docker compose up --buildThe API is exposed on port 8000 and the dashboard on port 8501. The first run downloads the model and encoder; named volumes retain both model and RAG caches.
Signal weights, verdict thresholds, model names, retrieval depth, and filesystem
paths are defined in config.yaml and validated by ragtag/config.py.
rag_backend: local selects sentence-transformers, FAISS, and Ollama. Set it to
openai_compat and configure openai_base_url, openai_chat_model,
openai_embedding_model, and (when required) openai_api_key for any endpoint
implementing /v1/chat/completions and /v1/embeddings. All generation calls
use llm_timeout_seconds and degrade to a deterministic unavailable-answer
fallback after a timeout or endpoint failure.