Skip to content
LlmWikis.orgLLM Wiki handbook
UAIX canonical
Menu
UAIX canonical
Source map LLM Wiki handbook llms.txt What is an LLM Wiki? Setup wizard Starter bundle vs RAG vs AI Memory Source policy UAI-1 spec

LlmWikis knowledge page

Graph Navigation

An LLM Wiki can stay plain markdown and still behave like a graph. The trick is to make the links typed, the index deterministic, the log append-only, and the clusters visible enough that an agent can route without scanning the whole repository.

Navigation layers

Layer Artifact Job Failure if omitted
Catalog wiki/index.md List every page with title, type, tags, summary, and status. The agent guesses paths or rereads too much.
Audit trail wiki/log.md Show what changed, when, why, and which sources were used. Drift becomes untraceable.
Typed links Frontmatter and wiki-links Distinguish supports, contradicts, supersedes, depends-on, and related-to. All links look equally meaningful.
Cluster map Topic or hub pages Group related pages for high-level routing. Large wikis become flat lists.
Contradiction map Contradiction pages or fields Keep disagreements visible until resolved. The wiki silently overwrites uncertainty.

Recommended link types

supportsThe source or page provides evidence for the target claim.
contradictsThe source or page conflicts with the target claim.
supersedesThe page replaces an older procedure, model note, or status record.
depends-onThe page cannot be interpreted safely without another page.

Small graph map format

## Graph Map

- cluster: ingest
  pages:
    - wiki/operations/ingest.md
    - wiki/operations/two-step-ingest-pipeline.md
  key_edges:
    - ingest.md depends-on page-schema-standard.md
    - source-summary-template.md supports ingest.md
  open_questions:
    - Should screenshots receive a separate source-summary type?

Search and vectors fit later

Vector search can help a large wiki, but it should not be the only navigation layer. Keep the human-readable index, log, hub pages, and typed links even if embeddings or graph databases are added later.

Hybrid search boundary

Search layer Good use Do not let it replace
Exact text search Find filenames, terms, commands, paths, and unique phrases. Source status and review labels.
BM25 or lexical ranking Rank likely pages in larger compiled wikis. The curated index and topic hubs.
Vector retrieval Find semantically similar compiled pages when wording differs. Traceable citations and contradiction records.
Reranking Prioritize a candidate set for the agent to read. Human review for high-impact claims.

Graph-aware lint checks

  • Find pages that have no inbound links from any hub or index section.
  • Find contradicts links without a contradiction note or review question.
  • Find supersedes links where the older page is not marked stale or archived.
  • Find clusters with too many pages and no summary or decision route.