mistralai / platform-docs-public Goto Github PK

Trying to use Mistral API as a drop in replacement for Openai in a project at the user's discretion, and I get the following error:

2024-01-24 14:17:01,574 INFO: message='OpenAI API response' path=https://api.mistral.ai/v1/chat/completions processing_ms=None request_id=None response_code=422
2024-01-24 14:17:01,574 ERROR: LLM API Error: Invalid response object from API: '{"object":"error","message":"{\\"detail\\":[{\\"type\\":\\"extra_forbidden\\",\\"loc\\":[\\"body\\",\\"frequency_penalty\\"],\\"msg\\":\\"Extra inputs are not permitted\\",\\"input\\":0.0,\\"url\\":\\"https://errors.pydantic.dev/2.5/v/extra_forbidden\\"}]}","type":"internal_error_proxy","param":null,"code":"1000"}' (HTTP response code was 422)

Even if Mistral API does not use other parameters Openai allows, is there a way for it to ignore additional parameters? It would make it a lot easier for clients to offer Mistral API as an alternative to Openai without writing special code just for Mistral. Here, frequency penalty is a standard Openai API parameter: https://platform.openai.com/docs/api-reference/chat.

Thank you for the consideration. I was not sure which repo to put this issue under, but chose documentation since if this will not change it would be good to put a specific note that Mistral does not support any additional errors and will not complete requests with any additional parameters.

Since overview does no longer exist, the link to the "Guides" page on the main page (https://mistral.ai/) no longer works.

Mistral uses the same bearer token security that OpenAI does. The OpenAI OpenAPI spec has it well defined (https://github.com/openai/openai-openapi/blob/master/openapi.yaml#L5021C1-L5025C23) on each API which mean any client library that's auto-generated knows how to manage security.

Our breadboard library currently can't work out consistently what the security spec is which means I can't offer a drop in replacement for OpenAI (which I would like to be able to do).

Happy to file a PR.

Hello!

I see that your documentation mentions that mistral-embed scores a 55.26 on the MTEB leaderboard. I'd love to help put this model on the MTEB Leaderboard as one of the External models. If you still have the results directory with the MTEB scores (i.e. a folder with a ton of json files like AmazonCounterfactualClassification.json, ArguAna.json, etc.), then I could include it into the MTEB leaderboard quite easily.
We do the same with OpenAI, Cohere & Voyage models, for example.

Feel free to reach out here, via [email protected] or via LinkedIn (@tomaarsen).

• Tom Aarsen

https://mistral.ai/ link to LinkedIn does not direct you to Mistral LinkedIn page.
When clicked it takes you again to https://mistral.ai/ again instead of taking you to https://www.linkedin.com/company/mistralai/

Hello, Could you do something for the open-mixtral-8x7b model to fix the truncation bug on "la plateforme".

Here they explain what they did to solve it on a vllm server (with spacing between <s> and [INST]).

On api where they are in completion I solved the bug by following the issue resolution.

vllm-project/vllm#2464