Skip to content

Wiki

The wiki is the central place for "how we do things", runbooks, service docs, and notes.

Concept

  • Folders (categories) group articles — every tenant has its own. If you have access to several tenants, you switch between them in the wiki; an article can only be filed in a folder of its own tenant.
  • Articles in Markdown, one language per article (default German) — with full-text index + vector embeddings
  • Service linking — an article can attach to a specific profile check, a check type, or individual hosts
  • AI RAG source — the AI chat draws from the wiki as its first source

Create an article

/wiki → New article.

Field Meaning
Title required
Category from list or new
Tags optional
Content Markdown, with toolbar and live preview
Service links n:m with profile checks / check types / hosts

Editor toolbar: H1/H2/H3, bold, italic, code, link, lists, table, code block.

Service linking

An article can be linked to:

  • a specific profile check (e.g. exactly this SNMP check on this profile)
  • a check type in general (e.g. all agent_disk checks, regardless of profile)
  • a single host (for quirks that don't apply to every host)

Effect: on the host detail page, the affected check's quick-lookup gets access to the linked articles; AI service analysis prioritizes linked articles in its RAG context.

Example: an article "Disk-full runbook" is bound to the agent_disk check type — once a disk-full alert fires, the runbook is right there, regardless of the specific host.

/wiki searches title + content with three combined methods:

  1. Full-text search (German FTS index)
  2. ILIKE fallback for partial strings (e.g. "backup" also finds "backup-script")
  3. Vector search (pgvector + embeddings) for semantically similar content

Embeddings

On article save, an embedding is computed in the background (asynchronously, so saving isn't blocked):

  • Model: the embedding model configured under AI provider (e.g. nomic-embed-text via Ollama)
  • Storage: wiki_articles.embedding, as an unconstrained vector column type — it adapts to whatever dimension the configured embedding model produces (different models return different vector lengths, e.g. 768 for nomic-embed-text)

Missing embeddings (e.g. after a bulk import, or if no provider was configured at save time) are backfilled periodically by a background job.

asyncpg cast for vector

For custom SQL queries with pgvector: use CAST(:emb AS vector) instead of ::vector — asyncpg interprets :: as a bind parameter and breaks the cast.

Quick lookup on service detail

A quick wiki suggestion is available right on a host's service row — without any LLM call, purely deterministic and typically under 100 ms:

  1. Direct hit via linked articles (check type, profile check, or host)
  2. Fallback: full-text search over the service name

The suggestions can be hidden per user in preferences if they get in the way.

Soft delete

Articles are soft-deleted (see Trash) — deleted means hidden first, physically removed after 30 days. Service links survive a soft delete and are fully back after a restore.

URL Effect
/wiki overview
/wiki?article=<id> open an article directly
/wiki?new=1 new article with an empty modal

Permissions

Permission Effect
wiki.view view articles
wiki.edit edit
wiki.create create new
wiki.delete soft delete

Permanently deleting from trash goes through the general trash permissions, not a dedicated wiki permission.

Next