An embedding index stores representations produced by a particular model and preprocessing process. Replacing the query encoder changes one side of the comparison. Equal vector length makes an operation possible; it does not establish that the resulting similarity scores mean the same thing.
Dimensions do not identify the coordinate system
Consider two models that each produce a vector with the same number of components. Each has learned its own representation. A matching component count does not prove that corresponding positions encode compatible relationships.
Sentence-BERT provides an example of sentence representations derived from a trained model and compared using similarity. The research does not establish that arbitrary models can exchange vectors. Treating a model change as a migration is a design deduction from this dependence, not a cross-model performance result reported by the paper.
Migrate documents and queries as a tested pair
A replacement plan can regenerate document vectors with the new model and evaluate queries against the resulting index. An alternative would require an explicit alignment method and evidence that it preserves the task-relevant relationships. Merely accepting the old vectors in the same database is insufficient.
Record the model version, preprocessing, chunking, dimensionality, and distance measure for each index. Retaining source documents and stable document identifiers allows the migration to be reproduced and retrieved results to be compared.
Retrieval quality needs its own evaluation
Take a fixed set of questions with reviewed relevant passages. Compare what the old and new systems retrieve, including misses and irrelevant high-ranked results. An answer model’s fluent summary cannot repair evidence that the retrieval stage never supplied.
Running the indexes side by side can make differences inspectable before cutover. Rollback also requires retaining a compatible query path, not only the old database files.
The migration succeeds when the new retrieval pair meets the stated task requirements. Successful vector insertion, matching dimensions, and an unchanged API are narrower technical checks.