RAG Retrieves. It Does Not Remember.

A RAG pipeline embeds a corpus of documents into vectors, then at query time embeds the question and returns the passages whose vectors sit closest to it. That is a similarity computation: it finds text that resembles the query's text, ranked by distance in an embedding space. It does not evaluate whether a passage is current, whether it was superseded by a later decision, or how it relates to any other passage — those are not properties similarity search represents at all.

It works well when an answer is genuinely self-contained in a document and the query is phrased close enough to reach it — a product spec, a how-to page, a policy stated in one place with no history of revision. In that case the nearest neighbor to the query usually is the correct passage, and the model quotes or paraphrases it accurately.

It fails when two versions of the same policy exist — one superseded, one current — and both are roughly equidistant from the query in embedding space, because nothing about embedding distance encodes recency or authority. It fails when the answer requires connecting several documents that don't share vocabulary: a contract clause, an incident report, and a pricing exception, none of which mention each other by name but which are causally linked. And it fails when the question is really about a relationship — "why does this rule exist," "what would break if we removed it" — because a relationship is not a passage any embedding model was trained to retrieve.

The failure is quiet specifically because the retrieved passage is often topically plausible. The model doesn't refuse to answer; it answers fluently from a document that happens to be wrong for the moment, and nothing in the pipeline flags that.

The fix is not a better embedding model — it is a structure that represents the relationship directly instead of hoping similarity approximates it. A system that records which decision produced which constraint, and what later changed it, can answer "is this still in force" because that is a lookup against structure, not a hope about distance. Venkai builds that structure and serves it alongside or underneath a retrieval step — retrieval still finds the right document; the relational layer says whether it's the current one.