It's definitely doable right now. llama.cpp (and frontends) can run in CLI on most CPUs. Mistral 7B is a "good enough" base model to run on most machines. Train

It's definitely doable right now. llama.cpp (and frontends) can run in CLI on most CPUs. Mistral 7B is a "good enough" base model to run on most machines. Training is good, but you can also inject current relevant man pages into the context with a vectordb or maybe even a simple grep. One issue is that LLMs tend to lie when they don't know really know a particular answer, and you usually solve that by double checking it with a web search.