Manage embedding versions in Milvus by storing model metadata, versioning collections, and incrementally updating outdated embeddings.
Versioning strategy:
1. Model metadata per embedding: json { "embedding_model": "text-embedding-3-large", "model_version": "20250101", "embedding_dims": 1536, "generated_at": 1700000000 }
Store alongside vectors. Agents know which model generated each embedding.
2. Collection naming: Create collections by model version ("embeddings_v1", “embeddings_v2”). Agents can query specific versions or compare results.
3. Gradual rollover: When releasing a new embedding model:
- Create new collection with new embeddings
- Run both in parallel (dual-write)
- Compare agent performance (relevance, loop count)
- Migrate agents to new model over 1–2 weeks
- Archive old collection after full migration
4. Incremental updates: Don’t re-embed the entire database. Track:
- Last embedding timestamp per document
- If source document is newer, re-embed
- Use background workers to incrementally update
5. Semantic versioning: When you update the embedding model, decide if the change is breaking (new dims), non-breaking (new model, same dims).
Fallback strategy: Keep previous embeddings available for agents that need stability. Newer agents use latest model.
For agentic RAG, embedding freshness is critical. Outdated embeddings (>30 days) degrade agent relevance by ~15%. Build versioning into your pipeline from day one.
Related Resources: