Journal
Build notes on my projects: what I tried, what broke, and what I learned. Newest first.
Web search without an API key
Local assistantBing's API is gone and Brave's free tier ended, so my assistant now searches through a SearXNG instance running in Docker on my own machine.
Teaching a 3B model to call tools
Local assistantPhi-3.5 has no built-in tool calling, so I had to fake it with prompts. Qwen2.5-3B has it natively, and the difference was 0 out of 6 against 6 out of 6.
A chat model in 4 GB of VRAM
Local assistantStep one of my local assistant. A 3.8B model, 4-bit quantisation, and a terminal chat loop that streams its answer.
