PRODUCTION SCENARIO
A biotech loads instrument operating procedures for three hardware generations into one OpenAI vector store. Scientists running the Gen-3 instrument keep receiving chunks from the Gen-1 procedure, which reads almost identically but specifies different reagent volumes.
Which change keeps Gen-3 users on the Gen-3 procedure?
Answering here is anonymous. Nothing is saved unless you sign in.
Show answer and explanation
Answer: Attach a generation attribute to each file and filter on it at query time
Each vector store file carries an attributes dictionary that attribute filtering can compare at search time, using comparison filters on a key and compound filters joined with and or or, so filtering drops the Gen-1 file out of the candidate set before ranking. Returning more results or lowering the threshold only admits more near-identical Gen-1 chunks, and two procedures that read alike are exactly what semantic ranking cannot separate.
Official Source: https://developers.openai.com/api/docs/guides/retrieval ↗