How does Pinecone handle vector updates and versioning?

Pinecone treats upsert as insert-or-replace by vector ID, so writing a new vector with an existing ID replaces the old one atomically. There is no built-in versioning or history, so if you need to track changes, store version information in metadata like a version number or timestamp and query with a metadata filter. For A/B testing embedding models, use separate namespaces or indexes rather than updating in place, which lets you compare retrieval quality before fully switching.