An embedding upgrade is an index migration
A new embedding model should go into a separate index and be tested on paired questions before a working search tool points at it.
Swapping an embedding model can look like a configuration edit. For a document search service it is a migration of the representation you search, so the existing index should stay available while you find out how the new setup behaves.
Google's embedding task-type documentation, read on 14 November 2026, distinguishes retrieval queries from retrieval documents. Those task settings should be saved with the model and dimensions, because a model name alone does not let you reproduce the embedding request.
What a fair comparison holds fixed
Start from a dated source collection with stable document identifiers. Keeping the same chunk boundaries for the first comparison isolates the embedding change; if chunking changes too, the combined configuration gets a new name and a description of both changes.
For each test query, mark which passage it needs. Include ordinary wording, exact identifiers and multi-document questions where your application sees them, and judge expected evidence independently of either index's first result.
| Invented retrieval outcome | Queries |
|---|---|
| Both retrieve evidence | 39 |
| Old only | 6 |
| New only | 3 |
| Neither | 2 |
In this invented set, the old configuration retrieves evidence for 45 of 50 queries, or 90%. The new one reaches 42, or 84%. Six old successes are lost and three old failures improve. For a migration decision, that paired record says more than the new total alone.
The candidate index is built beside the live one
The candidate gets its own identifier, and a manifest records the source snapshot, chunker version, embedding configuration and document count. Test queries go to the matching candidate configuration. Vectors from two models should not be assumed to share a useful space just because their lengths match.
Before a live route switches, check that the candidate collection is complete by comparing expected chunk identifiers with indexed ones and inspecting anything missing or duplicated. A retrieval test on a half-built collection mixes migration completeness with search quality.
The earlier index stays for a rollback period suited to the application. That is a proposed method, not a guarantee that a particular database supports an atomic alias switch, so verify the actual mechanism in your chosen system.
A rebuild's size comes from its text volume
Take a planning manifest of 100,000 chunks averaging exactly 600 tokens. Re-embedding that text means 60 million input tokens under the assumption. A cost estimate multiplies that by a verified rate for the candidate service; we have not supplied a price or sent those requests.
The volume says nothing about elapsed time. Request limits, batching and the indexing stage each need their own observations, so time a small representative run and record the conditions before extrapolating. Vector storage and temporary capacity sit outside the token estimate.
Documents edited during the rebuild must reach both versions. One approach is a fixed snapshot followed by an explicit change log. Whatever you pick, the candidate needs a clear cut-off and a verified final catch-up, because an old rebuild is not a current one.
The switch waits until the evidence can be read
The release decision should come with the paired losses and gains, the completeness check and the rebuild record. If the downstream answer stage depends on the retrieved passages, test that too, since better evidence coverage does not by itself mean better answers.
A new embedding release is a reason to test a candidate. The migration record should say which collection was rebuilt, which questions changed and which route restores the previous behaviour.