The answer
A new model can produce a different vector space. Query vectors and stored document vectors may no longer be compatible unless the collection is re-encoded consistently.
How are embeddings learned during model training?
Why can updating an embedding model affect an existing search index?
More FAQs in How are embeddings learned during model training?
No. An implementation can freeze an embedding layer while training other parameters. Whether embeddings change depends on which parameters the optimizer is allowed to update.
Not always. Token prediction can use text itself to supply targets, while other objectives use labeled pairs, interaction data, or automatically constructed examples.
Still have questions?
Browse all our FAQs or reach out to our support team
Want to upskill yourself?
Your learning library
Small steps. Big possibilities.
Find your next course. Build something great.