OpenAI compatibility with gpt-4.1

[MLBott]

Thanks for this excellent app. I encountered a few minor errors using OpenAI (provider) and gpt-4.1 (model). It could be my setup was wonky or OpenAI as an llm provider is still being worked on:

400 Setting 'max_tokens' and 'max_completion_tokens' at the same time is not supported.

After reading the comment on o3 and similar models, I tried commenting out 'max_comletion_tokens' and it worked fine.

There was this error after a connection attempt to OpenAI as provider:

400 Invalid schema for response_format 'schema': schema must be a JSON Schema of 'type: "object"', got 'type: "string"'.

But it works fine if non-object schemas are detected, and then, for OpenAI, wrapped in z.object({ value: schema }) for OpenAI compatibility, and then the response is unwrapped.

And the last one happened in a new chat, when trying to send the first message, just needed parameter filtering because OpenAI doesn't use those parameters:

400 Unrecognized request arguments supplied: dry_multiplier, min_p

Other than those, haven't encountered any other issues. And it's working fine for me after resolving them. Excited to see where this project goes!

[p-e-w]

Thank you for filing this issue, this is very valuable information. I hadn't tested Waidrin with the OpenAI API myself yet, and didn't know how deep the incompatibility would run.

These are relatively easy to work around, with the huge exception of the JSON schema issue. That one is a nasty surprise. It makes no sense to me why they wouldn't support non-object schemas, and fixing this without ugly special-case code might negatively impact the generation quality for backends that do, because the more wrapping layers you add, the more robotic responses tend to be.

[sheikheddy]

OpenAI compatibility also appears to be broken with Google's models. https://ai.google.dev/gemini-api/docs/openai#python

[p-e-w]

@sheikheddy

Which error(s) are you getting with those models?

[sheikheddy]

It would be straightforward to get a free API key at aistudio.google.com in order to reproduce the issue. See https://ai.google.dev/gemini-api/docs/openai for the base url.

…

On Mon, Aug 4, 2025 at 2:10 AM Philipp Emanuel Weidmann < ***@***.***> wrote: *p-e-w* left a comment (p-e-w/waidrin#29) <#29 (comment)> @sheikheddy <https://github.com/sheikheddy> Which error(s) are you getting with those models? — Reply to this email directly, view it on GitHub <#29 (comment)>, or unsubscribe <https://github.com/notifications/unsubscribe-auth/AHGAP6QO5GXUIFQOHHBUYDT3L32MTAVCNFSM6AAAAACCPMFPNWVHI2DSMVQWIX3LMV43OSLTON2WKQ3PNVWWK3TUHMZTCNBZGI2DSMJRGM> . You are receiving this because you were mentioned.Message ID: ***@***.***>

[p-e-w]

It would be straightforward to get a free API key at aistudio.google.com in order to reproduce the issue.

That requires providing a phone number where I live, and handing out my personal information just to reproduce a GitHub issue is a non-starter for me.

Waidrin is a Free Software project. I have little interest in cloud services and support isn't going to happen unless there is sufficient interest from the community to at least copy and paste whatever error message pops up.

[sheikheddy]

Hey, I'm sorry for incorrectly assuming that a verified Google account is something that everyone already has. I have uninstalled the repo for now since I can't use it but I appreciate the work you are doing, thanks and best of luck with the project!