Embedding drift is the gradual mismatch between user queries and document vectors in an index. It can occur even when documents stay unchanged, as people’s language and terminology evolve. In production, it may appear as vague chatbot replies or retrievals that are loosely related—or return nothing. section 4.6 section 4.6.1
Keep query and document embeddings aligned as content and models change:
- Re-embed documents when content changes, and reprocess the corpus when switching embedding models; otherwise, queries and documents can become numerically misaligned. section 4.6.1
- Use the same model version for query and document embeddings. Teams may re-embed nightly, weekly, or after model updates; no single cadence fits every system. section 4.6.1
- A sufficiently general embedding model can reduce how often—and how much—it costs to re-embed the corpus. section 4.7
