Embeddings And Retrieval

Team uses text-embedding-3-large but sets dimensions=256 to save vector-store cost, then compares to vectors stored at full dimension.

All vectors in an index must share the same model and dimensions; re-embed the whole corpus when changing dimensions. Mixing dimensions makes cosine similarity meaningless.

Retrieval ranks by dot product over un-normalized embeddings and gets inconsistent ordering.

OpenAI embeddings are normalized to length 1, so cosine == dot product; if the vector store re-scales or you mix sources, normalize consistently. Verify the distance metric matches the store config.

Org migrates embedding model; old and new vectors coexist during backfill.

Version the index by model; serve queries only against the matching-model partition until backfill completes, then cut over atomically. Never compare cross-model vectors.