Your RAG pulls snippets from my documents into the analysis. How do I know it retrieved the RIGHT passages — and that a bad retrieval doesn’t quietly poison the output?

The full executive answer

Let me translate the pipeline first: when you upload a document, we split it into passages, convert each into a numeric meaning-fingerprint — an embedding — and store those in a vector database inside your tenant. At generation time the engine searches those fingerprints for passages relevant to the module and your context, and where configured, a second-stage reranker — a cross-encoder model whose whole job is relevance judgment — re-scores a candidate pool of thirty before the best snippets are used. Fusion of ranking signals and the reranker exist precisely because raw similarity search alone is not good enough.

Against quiet poisoning, three controls. Retrieved text is treated as attacker-reachable — it passes the same injection-neutralisation as any untrusted input, and our red-team suite includes poisoned-document scenarios that must be detected and fenced as data. Attribution is explicit: every insight lists its grounded source documents by name, so a wrong-document retrieval is visible, not silent. And the retrieval layer is fingerprinted into the cache key, so a changed document library forces regeneration rather than serving conclusions grounded in deleted or superseded files.

Honest depth statement: we have a retrieval evaluation scaffold and the feedback loop feeds it, but we do not yet have large human-labelled retrieval gold sets or published recall metrics — that is roadmap, and the feedback capture is deliberately designed to build those sets from real usage. What I can demonstrate today is the attribution, the reranking, and the poisoning defences, live.