Mutli-LLM Providers & Generators

[simonprovost]

Hi folks!

Happy that I found a bit of time tonight to cook this up — I’ve had it in mind for quite a while. AutoDDG now supports multiple LLM providers via LiteLLM. Currently supporting by default OpenAI, Mistral, Anthropic, and Grok various models. Additional providers can be added by updating the config file or registering them manually (see README) — But I already spend £20 testing the above so will wait for future PRs from anyone if more have to be supported by default.

Introduced the following ~>

• Multi-provider support: Unified access to several LLMs through LiteLLM.
• Fuzzy matching: Uses TheFuzz to detect typos in provider/model names, suggesting corrections (e.g. Did you mean "gpt-4o-mini"? (92%)). Suggestions appear only for >90% matches. That's a personal touch, so much of a nightmare when you do not recall the exact name of a model.

Made the following API changes overall ~>

• Added a with_provider(.) for consistent provider/model usage now.
• Added helper functions to:
  • List available providers
  • Inspect provider parameters (for advanced users to be fair —— params being propagated into LiteLLM)
  • Retrieve supported models names per provider

Made sure examples are reported in new Notebooks ~>

• Provider Helpers: Demonstrates the new helper utilities in a nutshell.
• LLM Comparison: Benchmarks multiple LLMs on the same dataset.

Warning

Please do not merge this PR before #5 — Once #5 is done I'll rebase in this current PR will be up to date with main.

Cheers (Similarly, If @gracefan2020 wants to jump in to review the PR, feel free to do it!)

–––

Helpers — How to find providers and their models

AutoDDG_Providers_Acc.mp4

Bench multi-LLMs generated datasets description together

AutoDDG_Run_Multi_LLM_acc.mp4

[simonprovost]

Of course, I've removed all of the showcased API keys of the above video 🫡

[simonprovost]

@lyrain2001 @eduardar: Feel free to review at at your best convenience please :)

Cheers

[simonprovost]

Closing this because I will not have time to resolve the conflicts introduced by a newer PR. If those in charge still want LiteLLM, feel free to take over.