Use the same embedding model for indexing and querying. Mixing old document vectors with a new query model can make similarity scores misleading even if the dimensions match.
Evaluate before switching
Create a separate index, keep document chunks stable and test real queries with known relevant results. Compare the rankings, then switch over once the results meet your needs. Keep the old index available for rollback.
A migration test you can repeat
Before replacing the search index, collect a small set of questions with answers you can locate in the source documents. Include an exact-name query, a paraphrase and a question whose answer is absent. Save the current top results and the document IDs returned for each. These examples become your comparison set.
Build a separate index from the same source snapshot using the candidate embedding model. Give it an explicit model identifier and keep the chunking, filters and number of returned passages consistent. Query that index with embeddings from the candidate model. Keep the existing query path pointed at the original index while you test.
Suppose a query about refunds starts returning a delivery article after the model change. Check the retrieved passage before changing the answer prompt. If the refund passage never makes the top results, the problem begins in retrieval. If it is present but the answer ignores it, investigate the answer stage separately. Document whether each query retrieved the expected passage and whether the final answer used it correctly.
Only switch the production index and its matching query model together after the comparison meets your chosen acceptance criteria. Keep the previous pair available for rollback. A vector with the expected length is insufficient evidence that two models use compatible representations. This is a proposed evaluation procedure, with an illustrative failure.