A local Qwen model, a searchable slice of my website, and tools that propose exact edits. What worked, where retrieval went wrong, and why the useful engineering lives around the model.
A textbook confidence guarantee held 59.6% of the time on my real data. What I found, how I fixed it, what an independent review corrected, and a survival model for new security advisories with an honest account of what it does and doesn't show.