add gemini support and remove default models

[eamag]

Tried running it end-to-end using gemini, ran into so many errors and hidden default models, tried to fix them here. Can split/remove things like example project, ptal!

[eamag]

2025-04-15_23-01-47_automated_concept_sae_eval_attempt_0_reflection1.pdf
I managed to generate this in the end, figures are still missing

[conglu1997]

Thank you, would it be possible to make this a more minimal addition focused on adding Gemini support but changing nothing else?

[RichardScottOZ]

Adding gemini support means you have to generalise it more - because in the backend search it checks for either openai or claude models and that is all - so it will fail for llama, deepseek or other examples given in the llm list here.

[eamag]

@conglu1997 I assume you mean to remove examples? I deleted everything but gemini part and some bugs I found while running e2e

@RichardScottOZ Yes, there are some hardcoded models/clients in the code. I think the best way would be to add something like https://github.com/BerriAI/litellm but that was out of scope, I just wanted to test one run

[RichardScottOZ]

Yes, or the llama ideation uses OpenRouter already.. so another option for something like litellm

[RichardScottOZ]

Or something like this:- https://github.com/RichardScottOZ/AI-Scientist-v2/blob/main/ai_scientist/treesearch/backend/backend_openrouter.py

[RichardScottOZ]

Anyway, minor conveniences aside, AI Scientist v2 is very impressive.

Well done.

[eamag]

ptal @conglu1997

[conglu1997]

See #37

[eamag]

@conglu1997 that PR only changes models that were defined in llm.py, in your codebase you're also using just random clients defined in several places lol

[conglu1997]

Would it be possible to make a similar style PR? I'm concerned loads of new env variables are being used, new imports, etc. The ideal would be the minimal change that makes this work.

[BradKML]

Can this be generalized to cheaper models as well? #29