Support a local Ollama client

[jpmcb]

The Ollama project is an agnostic runtime for LLMs that runs models in containers: https://github.com/ollama/ollama this can be run locally or within a cloud environment.

I'd love to use gptscript, but I have no incentive to use openAI's gpt 3.5 or more expensive gpt 4 (while it's pretty cheap with lite use, over time, this could easily add up to $100s of dollars) when I have my own hardware and GPUs that I could run with codellama, llama 2, Mistral, or any of the other models the ysupport.

---

Proposal:

1. Refactor the pkg/openai package to be an agnostic type interface in pkg/client:

https://github.com/gptscript-ai/gptscript/blob/f162c5aa3f7971309d4a4347360fd43fa3e7c497/pkg/openai/client.go#L35-L40

2. Implement an Ollama client that sets the client.url to some env var provided by the user (in the local environment use case, this would likely be localhost)

---

I'd be happy to try and take a stab at this one since I have experience building local ollama integrations with neovim: https://github.com/jpmcb/nvim-llama

[ibuildthecloud]

You can set OPENAI_URL_BASE to point to ollama with gptscript today. Pretty much everything is standarizing to the OpenAI chat completion API and ollama supports that. The unfortunate thing is that no open source model really works well with function calling which is crucial to gptscript. We doing all we can to help move this along. Hacks, fine tuning, encouraging others, etc.

[HYY-yu]

I found a repo: https://github.com/NousResearch/Hermes-Function-Calling. It seems to mention compatibility with OpenAI function calling. @ibuildthecloud

[drpebcak]

Instead of setting OPENAI_BASE_URL we now have an alternative that is documented here: https://docs.gptscript.ai/alternative-model-providers

The same caveats apply though in terms of the support for function calling.