Is it possible to mix & match text embedding models?

[brasfb]

Can't figure this out by looking or asking elsewhere. If I use Sentence Transformers to embed a lot of text & store it in Chroma, can I query that using an OpenAI chat model? Or if I use ada for embedding, can those embeddings be queried by Cohere/LLaMA, etc?

[oddrationale]

Short answer: no. You want to use the same embeddings for the documents and the query.

That being said, it is possible to create alignment embeddings to "translate" one embedding model to another. Aligning embeddings involves finding a transformation that maps the embeddings from one model to another, so that their relative semantic relationships are preserved. Some research was done on this but it does not seem like an easy problem to solve.

[brasfb]

Thank you! I'm in my last semester of law school and my OpenAI credits ran out earlier than I thought they would, so trying to keep my costs low. I noticed that Chroma uses sentence-transformers by default and was hoping to make that work.

…

On Wed, Apr 19, 2023 at 1:50 PM Dariel Dato-on ***@***.***> wrote: Short answer: no. You want to use the same embeddings for the documents and the query. That being said, it is possible to create alignment embeddings to "translate" one embedding model to another. Aligning embeddings involves finding a transformation that maps the embeddings from one model to another, so that their relative semantic relationships are preserved. Some research was done on this <https://arxiv.org/abs/2101.07251> but it does not seem like an easy problem to solve. — Reply to this email directly, view it on GitHub <#3164 (comment)>, or unsubscribe <https://github.com/notifications/unsubscribe-auth/A43DXA6LS3MUYMYRY6M3YITXCAXVVANCNFSM6AAAAAAXEMD2QM> . You are receiving this because you authored the thread.Message ID: ***@***.***>

-- Brooks Brasfield 731.445.7942 Belmont University College of Law, Class of 2023 linkedin.com/in/brooks-brasfield <https://www.linkedin.com/in/brooks-brasfield/>