Structure pages with headings
Headings give each passage a clear topic, so retrieval finds it and citations point at the right section. A long page with no headings retrieves worse than the same content split under descriptive headings.- Use one topic per section, with a heading that names it the way a reader would ask about it.
- Keep a concept and its key details in the same section. A definition split across three sections is harder to retrieve than one self-contained paragraph.
- Put error messages, configuration names and product terms in the text verbatim. Readers ask with the exact string they saw.
Prefer sitemaps to crawls
A sitemap ingests exactly the pages you publish. A crawl ingests whatever links happen to connect, and can miss orphaned pages or pick up navigation noise. Use a crawl when no sitemap exists.Keep content current
Stale content is worse than missing content. Missing content produces an abstention; stale content produces a confident, cited, wrong answer.- Re-sync sources on a schedule that matches how often they change. MITHUNAI does not re-sync automatically.
- Remove sources you no longer maintain.
- When you publish a significant change, re-sync straight away rather than waiting for the next scheduled run.
Scope each assistant
An assistant answers best from a focused corpus. If you serve different audiences, such as public documentation and internal runbooks, create a collection and an assistant for each rather than one assistant over everything.Close the gaps your users find
Abstentions show what your users need that your content does not yet cover. Review the unanswered questions report regularly, write the missing content, re-sync, and ask the question again.Keep secrets out of your sources
MITHUNAI scans ingested content for credentials, but the reliable control is not to publish them. Exclude configuration directories,.env files and test fixtures from GitHub sources with exclude patterns. Every ingested page is treated as untrusted input; see Security.