Support using Embeddings Model exposed via OpenAI (compatible) API

[debanjum]

This change adds the ability to use OpenAI's embedding models or any embedding model exposed behind an OpenAI compatible API (like Ollama, LiteLLM, vLLM etc.). This allows using commercial embedding models to index your content with Khoj.

Khoj previously only supported HuggingFace embedding models running locally on device or via HuggingFaceW inference API endpoint.

[sabaimran]

Sweet! This should definitely help Khoj be more scale-ready for generating embeddings.

[jppaolim]

sounds amazing. But it doesn't seem to be actually indexing via the API. I used the MD file which is not yet merged, adding text-embedding-3-small to the name of biencoder, and inputing and API key. Shall I do something else ? How to make sure it works. ?

[debanjum]

Hey could you follow these instructions.

Two things to remember to do:

1. On the Admin panel, ensure only the Search Model Config for the OpenAI embeddings has the name default.
2. Restart your Khoj server before trying to index you md files

If it's still not working, it'd be good if you can share screenshots of your OpenAI Search Model admin config page and the top level search model config page from the admin panel that lists all the search model configs

[jppaolim]

Studying the log, I had passed https://api.openai.com/v1/embeddings and not https://api.openai.com/v1 as endpoint.
it seems to embed correctly. Then I had to adjust the treshold, I didn't perform extensive tests, but 0,18 was too low. I put 0,5 and it works for text-embedding-3-small. Further guidance will be welcome :)