Applied AI

Treat an embedding model as part of the index

Editorial edition · 16 November 2026

The idea

An embedding is a representation produced by a particular model and preprocessing pipeline. Similar vector dimensions do not guarantee compatible representations. Changing the model can change which passages appear close to a query even when the underlying documents stay the same.

Design choice

Version the model, normalisation and text preparation with the index. Rebuild or migrate deliberately when that contract changes. Keep an evaluation set so that a newer model is judged by relevant retrieval outcomes rather than by its release date.

Worked example

Two indexes run side by side during a model change. The same sample questions compare evidence quality before query traffic moves to the new index.

Before you ship

  • Record model and preprocessing versions
  • Never silently mix incompatible vectors
  • Keep a rollback index

Examples are illustrative. Adapt the checks to your own environment.

Further reading

Retrieval-Augmented Generation, Lewis et al.

Search the library