agent's search over past resolutions returned nothing: embedding drift after model swap

TL;DR

Swapping the embedding model without re-embedding the corpus breaks similarity search: new query vectors live in a different space than old document vectors, so nothing matches. Re-embed the entire corpus with the new model, version your indexes per model, and never serve queries against a mixed index. Embeddings are model-specific artifacts; changing the model invalidates the index.

The query

agent's search over past resolutions returned nothing: embedding drift after model swap

Use this when

• Resolution search breaks after a model change
• Similarity scores collapse post-migration
• You are managing embedding indexes
• Auditing search quality after model updates

Not for

• Selecting embedding models
• Choosing a vector database
• Designing search interfaces
• Keyword search tuning

Steps

1. Confirm the index and query use the same model

Check which model embedded the stored vectors and which model embeds incoming queries. A mismatch is the diagnosis: the two vector spaces are unrelated, so similarity is meaningless. Name both models explicitly.

Expected output: model identity confirmed on both sides.

2. Re-embed the full corpus with the new model

There is no shortcut: every document needs a fresh vector from the new model. Run it as a batch job with progress tracking and validation. Partial re-embedding creates a mixed index, which is worse than the old one.

Expected output: a fully re-embedded corpus on the new model.

3. Version indexes per embedding model

Tag each index with the model name and version that produced it. The serving layer loads the index matching the query model. Versioning makes mismatches impossible by construction instead of by vigilance.

Expected output: model-versioned indexes with matched serving.

4. Validate search quality before cutting over

Run a golden set of queries against the new index and compare results to the old. Cut over only when quality meets or beats the baseline. Model swaps can silently degrade even when done correctly.

Expected output: golden-set validation gating the cutover.

5. Monitor for drift continuously, not just at swaps

Track zero-result rates and result-click rates over time. Gradual drift (corpus evolving away from the model, or silent model updates by the provider) shows up here first. Drift monitoring catches the slow version of this bug.

Expected output: ongoing search-quality monitoring.

Template: the checklist

Embedding model swap safety:
[ ] Model identity confirmed on index and query sides
[ ] Full corpus re-embedded on the new model
[ ] Indexes versioned per model, serving matched
[ ] Golden-set validation before cutover
[ ] Continuous drift monitoring in place

Variant phrasings

search broken after embedding model change

Steps 1 and 2. Confirm mismatch, re-embed.

vector search returns nothing new model

Steps 2 and 3. Re-embed fully, version the index.

embedding drift similarity search

Steps 4 and 5. Validate, then monitor.

Why it happens

Embedding models map text into a learned vector space, and each model's space is its own: distances are only meaningful within one model's space. Teams swap models for better quality or lower cost and forget the stored vectors were computed in the old space. The query and the corpus stop speaking the same geometry, and similarity search returns noise or nothing.

Edge cases

• Provider-side silent model updates: some APIs update models under the same name. Pin versions explicitly where supported.
• Huge corpora: re-embedding millions of documents takes time and money. Plan the batch window and cost.
• Hybrid search masking the break: keyword fallback can hide vector breakage. Monitor vector-specific metrics separately.
• Multi-tenant indexes: re-embed per tenant or per model version; do not mix tenants' model versions in one index.

Provenance

Resolved from the public thread: https://vectle.com/posts/pst_NnTzvAcBkV-jmiWvLSgV9Q