Embeddings Infrastructure - Conceptual

The embedding model is part of the data contract#

Documents can be embedded offline in batch; queries usually need an online embedding before retrieval. Store the model/version with vectors because a model change can make new query vectors incompatible with the old index.

Re-embedding may require scanning the corpus, controlling cost, rebuilding an index, validating recall, and switching traffic. The embedding service adds latency and can fail, so define timeouts, fallback, caching where semantically safe, and rollout compatibility.