No matches
My wiki's raw sources passed 2 million words, and reading whole index files started to break down. Here is how I built hybrid search on SQLite and static embeddings after ruling out a heavier tool, and how I got agents to use it without being told the command.
My wiki search scored 0.95 on 20 queries, then 0.70 on 40 harder ones. Here is how I built the evaluation sets, which ideas failed, and which two changes raised the English score from 0.80 to 0.97.
What happened when I tried to tag my blog posts 'AI-assisted,' and why I landed on judging by effort and value instead of which tool touched the text.
I connected Mem0 to manage context across AI coding agent sessions, then removed it. Here's why two plain markdown files turned out to be enough.
Of 128 wiki pages, I had personally read almost none. When AI builds every connection, the knowledge graph belongs to the AI. It just lives under your name. How I built a weekly review workflow and a Slack safety net to keep my own thinking in the loop.
One source at a time, I reviewed carefully. Seven at once, and by the fifth I was scanning titles and rubber-stamping. How I recognized the browse-and-accept failure mode, and why splitting ingest into three separate stages fixed it.
Twelve entries seeded the source registry. Two had real evidence behind them. Ten were AI guesses stored in a file that looked like confirmed facts. How I reconstructed evidence from the ingest history, and what it means when confirmed and unconfirmed data look identical in a system you built to prevent guessing.
A public notebook on AI for knowledge work, grown by someone who spent two decades shipping and documenting hard systems.
Source trust judgments kept drifting between sessions, so I built a registry to make them persistent. The two-axis framework behind that design, the author-level extension that strengthened it, and the uncomfortable thing I found in the seed data I'd used to populate it.
Every summary Claude returned seemed obvious, until I realized my own impressions were being shaped by Claude's framing before I'd formed any of my own. What Sofia Quintero's essay helped me see, and how I built a friction step into my ingest workflow.
Personal notes, reusable templates, and public writing lived in the same space, and an AI agent had no reliable way to know which rules applied where. Why I split the vault into four explicit layers, and what the separation made visible.
Opening a wiki session, I saw 50,000 tokens already consumed before touching a single file. A 93 KB development framework had been loading into every session regardless of the task. What I learned about keeping AI context lean.
Looking for where DRM expertise actually meets the AI agent economy, my first answer was wrong. Here's the retraction, and the one real connection point regulation is already making mandatory.
I used to think feeding articles to AI and letting it fill my wiki counted as building knowledge. Here's how bad sources turned that into a well-organized trash can, and the three-part quality check I built to stop it.
Starting a public series on building a personal LLM-Wiki out of a decade of OTT and DRM experience. Why AI can summarize but not reflect, and what the system looks like a few weeks in.