A panel-llm style tool for the amalgamation of multiple responses to a single prompt from various llms of different temperatures, ensuring maximum detail and minimal hallucinations while alleviating the normal trouble of seven opencode tmux panes.
You can now use endpoints apart from opencode zen and custom models, defined with cli flags!
You can now get a result back in under 5 seconds with diffusion gemma on nvidia nim!
I uploaded it to pypi!
Everytime I try to upload to pypi, I always forget the format (src/…) and the commands with twine and build, but I’m getting there. I keep making minor errors (e.g. local imports can be absolute, but packaged version must use relative), and so I’m already up to 0.1.2 (two more updates to get it to a working state).
You can not install with pip install amalgam-panel! (Amalgam was taken already by a wrapper :<)
Added conversation / follow-up questions! The cli saves past prompts and responses, attaching to new messages as context, and allows for a more chat-interface style interaction.
I ran into lots of trouble with the ordering of messages and context structuring, but in the end it seems to work fine. Hope it’s a good change!
Added a context summarisation flag! You can now pass in –summarise-context for it to query free llms and summarise large repositories (any text-based files), saving tokens before the panel starts its work. Hopefully helpful for larger, unfeasibly large codebases, though rather dependent on the context-summarisation model (currently big-pickle, seemingly the most reliably-always-free model with respectable and reliable performance, with little errors or json issues).
While I was finishing up my Trisolaris project, along with polishing calculations for a fluid simulation project, I found myself asking multiple llms to check for errors or potential issues. This resulted in up to seven opencode instances in separate windows, and after getting their responses, I would then pass them to a reliable llm I trusted — often mimo v2.5 — to merge into a single cross-referenced response. While this was great, helping me to catch potential edge cases or give me some general advice on potential improvements, it was very messy and time-costly, and for my rather humble laptop, performance-draining. And so I thought I would use the opencode zen api and make a python tool to automate the process!
Amalgam is a python cli that accepts a variety of flags (to customise every aspect of it, from temperatures, number of responses, to timeout, and other parameters) and a prompt, presenting prompt and context to a ‘panel’ of llms (quite literally every free model available on the zen api, auto-detected with key validation), collecting numerous responses to be merged into a single unified, comprehensive response. This maintains the advantage of having multiple llms with different strengths, while keeping it overall time-efficient by automating many of the cross copy-pasting required normally, along with presenting all context (though this may be token-costly for larger repos, this could be improved in the future).
I’m quite happy with what I’ve got. It’s been a short project so far, but for my purposes, it’s practically done. I will work on it, potentially adding follow-up question features and context summarisation for larger repos, but it largely fulfills its purpose!
The below two screenshots show the before and after, from multi-window clusters to a singular, clean and unified report.
If anyone has suggestions or feature ideas, I would love to implement them! I feel like this is a perhaps common idea, but I haven’t find a solution for my needs (completely free auto-discovered model panel, with customisability yet simplicity); hope you like it!
