v1.7: Updates to OpenAI embedder
guimachiavelli opened this issue · comments
v1.7 introduces two minor changes to Meilisearch's OpenAI embedder.
- Two new supported models:
text-embedding-3-small
andtext-embedding-3-large
- Users can set the
dimensions
parameter for the OpenAI embedder (dimensions
must be >0 and lower than the model size)
Questions
- Why would a user manually set the dimension? What is the benefit of setting 100 dimensions when the model was built with 1500?
Tasks
- Update OpenAI instructions in
/experimental/vector_search
References
- SME: @dureuill
- GH discussion: https://github.com/orgs/meilisearch/discussions/677#discussioncomment-8306384
- engine issue: meilisearch/meilisearch#4394
ah I had missed that one, hello there
Why would a user manually set the dimension? What is the benefit of setting 100 dimensions when the model was built with 1500?
this is an "accuracy" vs "storage and compute" trade-off.
See the announcement blogpost from OpenAI, section "Native support for shortening embeddings" (why aren't people putting anchors on all section titles? I don't know...)