rag-sanitizer v0.2: escáner de corpus RAG pre-ingestión con embeddings reales y CLIP
rag-sanitizer v0.2: escáner de corpus RAG pre-ingestión con embeddings reales y CLIP Detecta documentos poison en un corpus RAG antes del chunk+embed, no después de la recuperación. v0.2 añade embeddings reales (MiniLM) y detección multimodal (CLIP). El problema La mayoría de la defensa contra envenenamiento de RAG opera post-retrieval o en inferencia . Pero el corpus se puede envenar antes de la…
rag-sanitizer v0.2 is a pre-ingest RAG corpus scanner that detects poison documents using real embeddings and CLIP. It identifies three types of poisoning: semantic mimicry, entity-swap, and visual poison. The scanner is designed to detect these issues before chunk+embed, preventing damage to the vector store. It offers three interchangeable detectors: Semantic mimicry, Entity-swap, and Multimodal (CLIP).
The installation process involves pip installing the needed packages and running the scanner with the appropriate embedder and multimodal options. The output reports are given per document, indicating CLEAN, SUSPECT, or POISON status along with the reasons. Measuring the real embeddings (MiniLM + CLIP) against a clean clone showed 22 tests passing out of 22.
The results suggested that the scanner is effective in detecting poisoning at the pre-ingest stage, even with limited resources. However, there are some known issues, such as handling fluid gibberish or images with low-resolution CLIP calibration. Regular external audits are recommended to ensure the scanner's effectiveness.
Written by urgent.news from Dev.to's reporting — not their text. Machine-written — may contain errors; check the original before relying on it.
