For Structured Data, Retrieval Accuracy Isn't a Model Problem

MongoDB shipped three retrieval features this week: context-aware embeddings, hybrid search, and a native reranker.

The pitch is the same one every vector database is making right now: retrieval accuracy is the thing that determines whether an agent's next step is right or wrong, and getting it wrong is expensive twice over, once for the bad action and again for the retry.

That framing is correct. It's also a framing built entirely around unstructured data, and it quietly assumes retrieval has to be approximate.

Why retrieval is probabilistic at all

Vector search returns the k passages whose embeddings are closest to the query's embedding.