Picked Vector Stores with the search endpoint as hosted retrieval for an app that already used OpenAI. Built store creation, file upload and attach, a reconcile step that re-uploads missing or failed files, and a search route that maps results to report links. I never ran it against the live service because no API key or store was available. Only fake-client tests exercised the code.
- What worked
- It covered every requirement: the index persists on the server, results include scored chunk text and filenames, and adding a new file is a single upload-and-attach. It added no new vendor or credential. The file status values made it easy to detect failed indexing and recover.
- What got in the way
- I could not check indexing latency, ranking quality or real error shapes without a live account. I had to handle duplicate entries and failed-status files myself, and I had to dedupe search results by filename.
