Research agentic papers and create architecture

[TanGentleman]

This codebase and the following papers should be very helpful in research: https://anonymous.4open.science/r/more_agent_is_all_you_need/ https://arxiv.org/pdf/2402.05120.pdf https://www.connectedpapers.com/main/6d8efe141ead5e397ca9a49814d049290292e974/graph

[TanGentleman]

https://arxiv.org/pdf/2404.01744.pdf

Octopusv2 paper. This will likely be the implementation if our function-calling use case scales. Here's the current plan:

• Create example input/output pairs for the specific task. This will be inputs that provide the allowed functions + query, with the output being the appropriate function call.
• Test Octopusv2 locally with few shot examples on simple function calling tasks appropriate for our project. If we need better alignment, proceed with fine tuning.
• If things go well, use a strong LLM (Maybe GPT-4 or Claude Opus) to populate a dataset of ~1000 input/output pairs for the fine tuning according to methodology in paper.

I should be able to fine tune Octopus locally on my hardware, but I haven't yet done many tests or build realistic pipelines that other LLM's couldn't handle with OpenAI tools/function calling. So I'll keep this on the backburner until the use for it organically develops.