Neural networks, deep learning, NLP · insight-0467
Re-embed the whole corpus whenever the embedding model or its preprocessing changes
- Evidence type
- practitioner heuristic
- Difficulty
- medium
- Roles
- ML engineer, Applied AI
- Industries
- saas
Summary: situation, action, reason, boundary
Situation
You upgrade the embedding model for new documents but keep the existing vector index built with the old model.
Action
Store the model name, version and preprocessing hash with every vector, refuse mixed-version queries, and rebuild the index in full (blue-green) before switching traffic, rerunning the retrieval evaluation set.
Reason
An embedding space is defined by the specific trained model; there is no shared coordinate system across models.
Boundary
Vector search, deduplication stores and feature stores holding embeddings.
Symptom or risk
Vectors from different models, versions or preprocessing live in different spaces, so similarity between old and new items is meaningless and retrieval degrades unevenly.
Exceptions
Methods explicitly trained for backward-compatible embeddings; verify that claim on your data.
Trade-offs
Full re-embedding costs compute and time for large corpora; schedule it like a migration.
Pitfalls
Changing text cleaning or chunking without bumping the version, which shifts vectors silently.