Upgrading or swapping embedding models in a retrieval-augmented generation (RAG) pipeline is more than a simple configuration change. A recent field report highlights a common but often overlooked issue: the new model's output vector dimension may differ from the old one, breaking compatibility with vectors already stored in the database. Even after re-parsing documents, the system can still fail because the historical vectors remain in the old format. This mismatch can cause silent errors, degraded retrieval quality, or outright failures that are hard to trace. The lesson for engineering teams is clear: before switching models, audit the dimension of existing vector data and plan for a migration or re-embedding strategy. This is an evergreen operational concern for anyone building or maintaining RAG systems, especially as the model landscape evolves rapidly.
A practical warning about vector dimension mismatches when migrating embedding models in RAG systems, and why historical data compatibility matters.