A supplier changes the search model behind its knowledge assistant. The manuals stay the same, but an answer about operating conditions now retrieves a general brochure instead of the relevant restriction. For a buyer, this looks like unreliable product information. For the implementation team, it may be an index migration problem rather than a writing problem.

Google's October 6 release of EmbeddingGemma 2 adds a multimodal embedding option with adjustable output vectors. It is a useful occasion to ask what happens to an existing product index when its embedding configuration changes. The migration approach below is an engineering recommendation, not evidence that this model will outperform another model on a particular supplier catalogue.

Test the evidence a buyer actually needs

Start with questions whose correct supporting documents are known. Include a precise model reference, an application described without the model name, a question about an exclusion, and a request the available documents cannot answer. A test made entirely from brochure headings will not expose the harder retrieval failures.

For each question, record whether the correct document appears among the candidates and whether the necessary passage is present. A manual found without its operating limitation may still produce a misleading answer. Comparing average similarity scores alone will not tell you whether the assistant can support a sourcing decision.

Keep a distinction between retrieval and answer generation, as in this diagnosis of a wrong product answer. If the right passage is absent, changing the final response prompt may conceal rather than fix the problem. If the passage is present but the answer omits its condition, the failure belongs further downstream.

An international test set also needs the expressions buyers use, not just translated internal labels. A product family may be searched by function, local terminology or an installed component's name. These queries can be compared across old and new indexes without changing the underlying technical facts. Language coverage should be tested directly rather than inferred from a model's general benchmark.

The same vector length does not establish compatibility

Embeddings are numerical representations created by a model. Retrieval compares a query representation with document representations. The comparison needs a compatible configuration on both sides; numerical length alone does not show that the coordinates have the same meaning.

A different output length can cause an obvious database error. A change that preserves length can be harder to detect: the query runs, candidates arrive, and their order looks plausible. Yet the old document vectors and the new query vectors may no longer form a meaningful comparison. Successful execution is therefore a weak migration test.

This is different from a model's explicitly supported vector truncation. EmbeddingGemma 2 describes that output option. It does not establish compatibility with vectors from another model. If the change is a supported truncation within the same model, check the required processing configuration and test retrieval rather than assume that every dimension change requires rereading the whole corpus.

The configuration includes more than the model name. Version, output dimension, normalization and document preprocessing can affect the result. Chunking deserves attention too. A previous chunk may have kept a model identifier and its limitation together, while a revised splitter separates them. A change attributed to the model may actually come from this different evidence unit.

Treat these settings as one identifiable index configuration. A support investigation should be able to establish which configuration served a disputed answer. Without that record, a team can repeatedly change prompts while the source of the retrieval shift remains hidden.

Build beside the old index, then move queries

Preserve original files and stable document identifiers. For a cross-model migration, create a separate index and regenerate its document vectors with the intended configuration. Generate test query vectors using that configuration as well. Do not mix old and new vectors simply because a storage layer accepts them.

During comparison, keep the source corpus aligned. If one index contains a new manual and the other contains an old manual, the experiment cannot isolate the embedding change. Document updates during the trial should reach both indexes through a controlled process, with their versions visible in the test results.

If image or video retrieval is being introduced, give it a separate test group. Passing text questions does not prove that an image query returns the correct product version. Require a result that opens the underlying asset or document so a person can inspect the match, rather than accepting an untraceable description.

Move a limited set of queries only after the evidence comparison is acceptable, and retain a route back to the previous index. This is especially useful when a Chinese supplier serves several overseas sales teams whose terminology differs. One team's improvement can otherwise hide another team's loss of relevant results.

A small text-only catalogue may not need a migration at all. If its current retrieval works, retain that configuration until a specific need justifies changing it. Rebuilding an index cannot supply a specification the source documents never contained.

Sources

Google, EmbeddingGemma 2: an open, lightweight multimodal embedding model, October 6, 2026.