A document index uses embeddings from model A. A developer replaces only the query encoder with unrelated model B, which has the same vector dimension. Search quality collapses. What should the team do?

1. Increase the language model output token limit.
2. Encode queries and documents with compatible models in the same embedding space.
3. Keep model B and normalize only the query text to uppercase.
4. Compare the vectors by their alphabetical representation.