The document count matches, nothing threw, and a specific set of documents is now unreachable by any query. That combination narrows the cause fast, and the two survivors are an approximate-search parameter that a rebuild reset and a metadata filter applied in the wrong order.
Updated Sep 2026 · Grounded in real Forward Deployed Engineer interview loops and written to a senior-engineer editorial bar.
The document count matches, nothing threw, and a specific set of documents is now unreachable by any query. That combination narrows the cause fast, and the two survivors are an approximate-search parameter that a rebuild reset and a metadata filter applied in the wrong order.
Two facts in the prompt decide this question and most candidates walk past both. The count is right, so ingestion did not drop anything, which eliminates the entire class of pipeline explanations people reach for first. And it is a subset rather than a random scattering, which is the signature of approximate search rather than a bug, because ANN recall is not uniform over a corpus and the documents in sparse regions of the vector space are the ones that go missing first. A candidate who fetches one of the missing documents by ID before theorizing has already won half the round.
No comments yet — be the first to share your approach.
