Hey,

When working on complex projects involving multi-document or multiple tasks process, I've stumbled upon a problem that LLMs can't guide themselves efficiently through context. I've seen some tools for LLM graph interface (https://branchcanvas.com/, http://grafychat.com/) but:

  - They are closed source and cost $
  - They are ugly (subjective) or clunky
  - They do not support merging the branches back

So I built Bushchat: an opensource, browser-based tool that turns LLM conversations into a tree structure, allowing you to expand and merge back when needed. This is basically LLM 'thinking' but with manual thread control, which increases the reliability of steering the azimuth of the thought process, and efficiently work on context.

In the future I envision that making this a collaborative tool where users could share and work together on live data trees or part of their context with each other could increase the productivity in heavy-AI reliant teams.

It's free, hosted on gh-pages and you can access it https://bushchat.xyz/ or run it yourself https://github.com/srakai/bushchat

Everything is DOM based, there is no server, but you need your API key or local LLAMA with OpenAI compatible API. Also disclaimer: There is posthog analytics, which is used to collect solely usage metrics (gh-pages does not facilitate that). Messages, API keys, responses, chat names are out of scope (https://bushchat.xyz/privacy ).

Happy to hear your thoughts!