Cited-Source Retraction and Recency Auditor for RAG Pipelines That Catches Confidently-Wrong Citations Before They Ship
Production RAG pipelines confidently cite retracted research papers, outdated regulatory text, and superseded versions of internal docs at high relevance scores. Teams building professional-grade AI (legal, medical, financial research) need an audit layer that, before any retrieved doc is fed into the LLM context, checks it against retraction databases (Retraction Watch, PubMed), document-version stores, and last-updated metadata, then flags or filters hits with stale or pulled provenance.
The trap is making it generic. Pick ONE vertical (medical research, legal precedent, FDA filings) where retraction or supersession has a real legal cost, and sell as a specific liability product rather than a horizontal RAG plugin.
landscape (3 existing solutions)
The infrastructure pieces exist (retraction DBs, vector store filters, observability platforms) but nobody has stitched them into a 'no retracted citation passes' middleware. For regulated verticals, this becomes a liability shield.