Retrieval that returns nothing useful
Documents were embedded and indexed months ago. The query path was written later, by someone who picked whichever embedding model was current that week.
Nothing errors. The vectors have compatible dimensions, the search returns results, and the results are noise — because the two models put meaning in different places, so distance between their vectors measures nothing at all.
Boundaries, outermost first: Indexing · ran months ago: Documents (chunked), Embedding v1 (used for indexing) Query path · today: User question, an empty slot for the embed the question Outside every boundary: Vector store (v1 vectors; FAILED: returns noise) Connections: Documents calls Embedding v1 (step 1) Embedding v1 calls Vector store — index (step 2) User question calls embed the question (step 3) embed the question calls Vector store — nearest neighbours (step 4)