Short answer: use four gates for marketplace listing discovery: an explicit collection, an allowed audience, a bounded result count, and source context that survives retrieval. Re-index changed listing chunks deliberately, remove deleted listings, and promote an index only after a small labeled evaluation shows acceptable retrieval quality at the latency the product can tolerate.

This is an access-boundary problem before it is a ranking problem. A highly relevant private listing in the wrong answer is still the wrong answer.

Start with the retrieval contract

Write the contract before selecting a vector service or endpoint. For a marketplace, the retrieval unit might be a listing chunk rather than the whole document, but every chunk must retain its listing ID, collection, audience, version, and source reference. The answering layer should receive enough context to trace a result back to the material that produced it.

The contract is small on purpose.…