compu-master: A Self-Hosted Research Companion on Local Silicon

In The Local Silicon Equation, I wrote about keeping a 128GB NVIDIA DGX Spark and an RTX 4090 in my office to run large open-weight models locally. compu-master is the tool I built on top of that hardware after my paid search API credits ran out on an older research script.

Rather than paying per query, I wanted a self-hosted tool that searches the web, reads the source pages, and compiles structured study notes locally—with every statement linked to a direct quote from the original source.

Quoting Sources Before Summarising

Most research scripts feed raw search snippets into a language model and ask for a summary. Once a model paraphrases a messy web page, it is hard to tell which facts came from the page and which ones the model quietly made up.

compu-master separates source extraction from writing. It queries a self-hosted YaCy search node, strips out page clutter, and saves verbatim quotes before the local vLLM model drafts a single note. A final check compares every statement in the draft back against those saved quotes and drops anything the sources do not support.

Practical Lessons from Running Locally

Pointing the pipeline at local reasoning models quickly surfaced a common quirk: when asked for structured output, reasoning models often spend their entire token budget "thinking" out loud and cut off before finishing the answer. Turning off the extended reasoning pass on straightforward extraction steps—and retrying without it if a response gets cut off—kept the pipeline reliable.

Running a tool that scrapes the web and generates charts on a local machine also meant taking basic care with isolation: running chart-plotting scripts in a stripped-down environment with no access to system secrets, and blocking web requests from following redirects onto local network ports.