Add "url" support for OpenAI embedder

[DavertMik]

It's a known thing that lots of AI providers use OpenAI API as their standard.

It would be nice if we could use OpenAI API but specifying a host other than api.openai.com

This would allow to use Microsoft Azure OpenAI and other providers as DeepInfra (which uses OpenAI API as well https://deepinfra.com/BAAI/bge-large-en-v1.5/api )

So I would suggest to extend configuration for OpenAI so we could use other vendors but the same API

[dureuill]

Hello 👋

Thank you for the suggestion. It so happens that the upcoming v1.10 is adding the url parameter as an optional parameter 🎉
That said, using this new parameter for the use cases you describe is not recommended ⚠️

This would allow to use Microsoft Azure OpenAI

Microsoft Azure OpenAI uses non-standard headers for authentication.

other providers as DeepInfra

These providers are not using the OpenAI API, they're using an OpenAI**-like** API. Some examples of the differences:

1. the model fields take a different value ("BAAI/bge-large-en-v1.5")
2. they don't support inputs in the forms of tokens tokenized with OpenAI's tokenizer
3. the token limit is not the same as for OpenAI limit (8196 vs 512)

The openAi embedder of Meilisearch is specialized to communicate with a genuine OpenAI endpoint. The url parameter was mainly added to serve the use case of people who proxy requests to OpenAI.

Instead, you should use the rest embedder source. Meilisearch v1.10 makes this more convenient:

1. rest embedders support custom headers, allowing Microsoft Azure OpenAI
2. the configuration of rest embedders is much simpler, with the newly introduced "request" and "response" parameters. According to their request/response example, the DeepInfra embedder would have the following configuration:

"default": {
  "source": "rest",
  "url": "https://api.deepinfra.com/v1/openai/embeddings",
  "apiKey": "<YOUR_DEEPINFRA_TOKEN>",
  "request": {
    "input": "{{text}}",
    "model": "BAAI/bge-large-en-v1.5",
    "encoding_format": "float"
  },
  "response": {
    "data": [
      {
        "embedding": "{{embedding}}"
      }
    ]
  }
}

or, to optimize throughput, as it seems supported by DeepInfra:

"default": {
  "source": "rest",
  "url": "https://api.deepinfra.com/v1/openai/embeddings",
  "apiKey": "<YOUR_DEEPINFRA_TOKEN>",
  "request": {
    "input": ["{{text}}", "{{..}}"],
    "model": "BAAI/bge-large-en-v1.5",
    "encoding_format": "float"
  },
  "response": {
    "data": [
      {
        "embedding": "{{embedding}}"
      },
      "{{..}}"
    ]
  }
}

More information about the newly added url parameter.

More information about the new REST embedder API

[DavertMik]

Thanks for the detailed explanation on this issue!
Really appreciated. I was reviewing RC release notes and didn't find mention of url param in there so I started this discussion.
However, your reply is more valuable than that, I will consider using REST embedder once the 1.10 is out!

Thanks so much!

[dureuill]

Happy to be of service :-)

I will close the discussion, feel free to reopen as needed ☀️