One thing you might notice is that I don't use classic tool calls - one of my early discoveries was that they are terrible from a UI perspective - the LLM goes radio silent before suddenly pumping out the call - its' easy to teach it to output them json block and parse the markdown output.
Doing these large runs also catches a few issues - like the 100% cpu as it kept building up a backlog of scroll to end etc.
Towards the end of this replay (and yes it can just about replay a session) - the markdown parser hits issues that hopefully be fixed soon
github.com/roojs/OLLMchat
So #ollmchat skill like agent is now running very long processes (over 500 requests for this one)
* it got all the way to writing the plan based on the original prompt (after full analysis)
* as these sessions get longer - its clear i need better visualization of progress
#vala #gtk #strix-halo
youtu.be/gL2gkVm3KWs
New progress video for #OLLMchat - the full skill agent workflow is beginning to work - task based refinement and tool running - all #gtk #vala running on #strixhalo with #qwen3.5
github.com/roojs/OLLMchat
Been quiet for a while - i've been looking at task based and skill based code writing - the framework is there - but it needs quite a bit of testing (and bit broken currently) - demo with #qwen-3.5 #ollmchat #vala #gtk
updated docs roojs.github.io/OLLMchat/oll...
github.com/roojs/OLLMchat
ollmchat's streaming markdown renderer - about 98% done i guess
images might be fun..
but tables and links work find now
#vala #gtk #ollmchat
github.com/roojs/OLLMchat
new preferences dialog for #ollmchat #vala #gtk #llm #ai
* manage connections
* manage models (set up context etc.)
* add models
* search ollama available models
* pick which size/version etc..
* demo is a fake download... - just for testing)
Christmas demo #vala #gtk #tree-sitter #llm #faiss
Code search tool using semantic search demo (since we normalized the vectors - the results are quite good)
#qwen3-coder:3b using tools and reading files
still plenty of UI issue to solve but the fun stuff delivering results now.
semantic code indexer running of #ollmchat
* uses treesitter to do structure
* uses LLM to do anaylsis and summary (Qwen3-coder)
* uses sql to to store metadata
* uses faiss for vector storage
#vala #gtk #ai working nicely.
The original semantic search plan was to use the llm and not tree sitter, but it turns out llms like to hallucinate code,
so going to create tree sitter Debian language packages, and use the llm to generate general descriptions. Among with vapi bindings to tree sitter, fiass bindings are working
So latest update on OLLMchat progress, my #vala #gtk #llm chat app. Video to come soon.
I've now completed about 80% of code editor integration. So you can view and in theory edit files..
Still having issues with the llm writing and editing. semantic code search is about 60% done. So getting there
OLLMchat latest video - I've got markdown working with basic block level (headlines and a few things) - also using sourceview to do rendering of code - all this is using strix halo box - there is an example of it answering real questions github.com/roojs/OLLMchat
Lots of under the engine changes - visually we now have a working model selector / tool enable / disable features
The streaming markdown renderer almost there, new code.
Currently sorting out edit as the LLM takes to long to make the tool call (so changing the way thats done) use last code in chat
#ollama #vala #gtk chat client coming along nicely - got it's own repo now github.com/roojs/OLLMchat
* function calling working
* updated the styling
* tool can send messages to the chat
next up #permissions (code done, but need to hook it into the interface)
screenshot of the example.sh running not re-indexing, and searching using faiss
todays update
* replaced the embed with a real vectore db - #faiss
* no more specificing database - now in ~/.local/share/semantic_code_search
* one big database for all code
* search - specify which files to search and/or specific language(s).
* updating of code only update the file you changed
filled in instructions to actually get it to work, and build (some days getting python to install stuff seems like a miracle..) - anyway this should work as the code search for the LLM github.com/roojs/semant...
#ai #vala coding gets closer - been trying to fight #python all day... - finally removed tree-sitter python dependencies in semantic code search (it was broke, it now uses the command line version, and we can specifiy what files to use)
github.com/roojs/semant...
added some nice formating to the ollamachat widget
roobuilder gets AI ready - pure #vala #gtk and #ollama local #llm - it talks so far - quite a way to go .. tools etc.. .
In true HK style, even after talking to support, the empf system was unusable, so had to fax the the submission #empf #hongkong
sorry system not working - HK's new 'empf' system down on first day
First day companies have to pay using the new HK Gov MPF payment system - (we used to use the providers websites) - and it's been down so far nothing works - not a great start
Not bad, pretty much out of the box, 53tokens/s using gpt-oss 20b #ai #debian gmktec ai evo x2 #lmstudio
Just got the gmtek AI machine, installed Ubuntu 25.10, but after seeing, no xorg, snaps everywhere, decided time to install debian again. Begining to forget why I switched to Ubuntu from Debian originally
A new video, this time using #cursor and #ai to modify the #vala code in roobuilder youtube.com/watch?v=A6WV... along with me..
That's being very patient, or just distributed attack, looking at the latest AI tools, running any kind of public server, is going to be far more work than it used to be
We noticed a few months ago what looked like massive number of requests from mobiles phones doing spidering. Suspected to be mobile proxies, mostly from China unicom, ended up forcing all China to do challenges. As the only solution. Not sure if it's apps with that added as a revenue model.
www.1news.co.nz/2025/10/05/v... not a bad article, mirror our recent switch to ai based coding. Stuff that is always low priority like testing docs, can be given to ai. Senior engineers review and adjust generated code. And try and write rules for AI, which it seems to rarely follow unless told to
With date, confirmed
