Pebblo

Pebblo answers the question every RAG team dreads: what is actually in the data we are feeding the model? It wraps your framework's data loaders (LangChain supported) and runs semantic topic and entity classification over every loaded snippet, then summarizes the findings in a UI dashboard or a PDF report you can hand to compliance. It also ships a SafeRetriever that enforces identity and semantic rules on vector-database retrieval before LLM inference.

The pitch is speed to production: instead of a security review discovering PII or toxic topics late, Pebblo surfaces them during development, with reports aligned to regulations like the EU AI Act. It is MIT-licensed and pip-installable.

The honest caveat: the project has been quiet since mid-2025 (~149 stars, last push June 2025) and the LangChain-only loader support never broadened to LlamaIndex or Haystack as promised. The classifiers are heuristic/ML-based, not guarantees — treat Pebblo as a safety net and audit layer, not a compliance certification. Still useful as a lightweight, self-hosted check before a RAG app goes live.

1. pip install pebblo --extra-index-url https://packages.daxa.ai/simple/ (Linux may need libpango system packages).
2. Start the daemon: pebblo (listens on localhost:8000).
3. Wrap your LangChain data loader with Pebblo's Safe DataLoader (pip install pebblo-langchain).
4. Load data as normal — review topics/entities in the UI at localhost:8000 or export the PDF report.