Skip to main content
Client access Talk to us
Artificial intelligence12 min read

`embedding-2-preview` reached shutdown — changing the ID without rebuilding the index breaks search

The official table lists 10 August as shutdown and points to `gemini-embedding-2`. Vectors from different models should not be mixed as if they shared one semantic space.

Updated
Nature
Model lifecycle shutdown notice
Date
Shutdown listed for 10 August 2026
Replacement
`gemini-embedding-2` in the official table
Limit
Retrieval equivalence depends on corpus, chunking, metric, and application
Two vector archives are compared through a calibration bridge
An embedding migration ends only after search, indexes, and rollback are measured on the same corpus.AI-generated image under Velqory editorial direction.

In brief

Google released `gemini-embedding-2` as GA on 22 April and listed `embedding-2-preview`, launched 10 March, for shutdown on 10 August. Migration affects writes, search, indexes, caches, and evaluation. The safe path builds a separate index, dual-writes where possible, rebuilds the corpus, measures recall and ranking, and changes reads only after an explicit gate.

  • Find direct calls and persisted embeddings; changing only the producer mixes incompatible spaces.
  • Create a new index with model version, dimension, normalisation, chunking, and content hash.
  • Evaluate retrieval with labelled questions and relevance, including real tails and languages.
  • Plan read rollback; reverting code does not undo vectors already written.

The endpoint shuts down; old vectors keep influencing results

The deprecations table names a replacement but does not promise numerical equivalence. An index compares distances within one space; vectors produced by different models or settings can make scores incomparable.

Release notes describe multimodal input in the GA release. That expands possible designs and required tests: modality, input size, output dimension, cost, latency, normalisation, and retention policy need recording.

Migration without a blind cut

  1. 22 APR
    GA model

    `gemini-embedding-2` becomes the stable destination.

  2. 10 AUG
    Listed shutdown

    The preview can no longer be assumed as an available producer.

  3. REINDEX
    Corpus rebuilt

    A new index receives every document with provenance and reconciled errors.

  4. CUTOVER
    Reads promoted

    Recall, ranking, latency, and cost pass before switching.

Documentary basis for this section

Compare complete retrieval systems, not isolated vectors

Freeze representative queries and relevance judgements. Run old and new indexes with the same filters, top-k, and post-processing. Measure recall@k, nDCG or a suitable metric, unsupported answers, latency, and cost per accepted query.

Reindexing should be repeatable and auditable. Retain model, dimension, chunking, document version, timestamp, job, and status. Failed documents must enter a reconciliation queue with an owner rather than silently disappearing.

New-index gate

LayerEvidenceThresholdRollback
CorpusCounts and hashesNo unexplained gapsRerun job
RetrievalLabelled recall/rankingNo critical-case regressionKeep old reads
AnswerCitations and supportNo material unsupported riseStop promotion
Operationsp95 and errorsWithin SLOPrevious routing
CostAccepted queryWithin budgetAdjust design
Documentary basis for this section

Dual-index and switch reads only after the new index is complete

While the system is still writing, send new documents to both indexes and reconcile counts. Rebuild history through idempotent jobs; keep the old index while rollback and observation remain active.

After cutover, monitor score distribution, empty results, clicks or acceptance, and failures by language or modality. Keep model changes visible in telemetry so data drift can be distinguished from an embedder change.

  • Do not mix vectors from two models without explicit proof.
  • Do not validate only on easy questions selected by the team.
  • Do not close reindexing with failed documents outside a repair queue.
  • Do not delete the old index before exercising read rollback.
Documentary basis for this section

Sources consulted

Consulted on the stated update date.

  1. 01Gemini deprecationsGoogle AI for Developers · AI-GOOGLE
  2. 02Gemini API release notesGoogle AI for Developers · AI-GOOGLE
  3. 03EmbeddingsGoogle AI for Developers · AI-GOOGLE
  4. 04Gemini modelsGoogle AI for Developers · AI-GOOGLE
  5. 05Challenges to the Monitoring of Deployed AI SystemsNIST · GL-NISTAI
  6. 06AI Risk Management Framework CoreNIST AIRC · GL-NISTAI
← Back to the Blog