Principles for running an AI chat on a home server
Running an AI chat at home means doing yourself what a cloud service would do for you. These are the four principles I care about most.
1. Answers need evidence
Answers that use web search carry source numbers. While merging with this site, the source cards gained the time a source was checked, its publication date (or “unknown”), and whether it is site or web material.
2. Never touch the internal network
When the AI opens a link whose address points to a private or internal network, it refuses. Some addresses look public but redirect inward, so it re-checks on every redirect.
3. Store as little as possible
Public Ask AI questions are not stored. A one-way hash, timing, success and source URLs are enough for operations.
4. Machines look at quality first
Every answer is checked against rules, and suspicious answers are graded by another model. A nightly regression run shows immediately where things got worse after a model change.