Skip to main content
Without typed pages, a compiled wiki quickly becomes a flat list of topics with no signal about what role each page plays. A page about “Transformer” could be a conceptual explanation, a specific model card for GPT-4, a side-by-side of Transformer vs. RNN, or a domain map tying together all attention-related pages. llmwiki’s four page kinds give the compiler - and you - a vocabulary for these distinctions. Each kind carries different linting expectations (such as minimum wikilink counts you can configure in schema.json), surfaces differently in the local viewer, and participates differently in the review policy.

​
The Four Page Kinds

concept

A standalone idea, technique, or pattern. This is the default kind and the most common output of compilation. Use it for abstract or generalizable knowledge: algorithms, design patterns, terminology, theoretical frameworks. A concept page explains what something is and links to related concepts.Examples: self-attention, knowledge-compilation, incremental-compilation

entity

A specific, named thing that exists in the world - a person, organization, product, model, or artifact. Entity pages are about particular instances rather than abstract patterns. They carry unique identifying information and often link outward to the concepts they instantiate.Examples: andrej-karpathy, gpt-4, anthropic, attention-is-all-you-need

comparison

A side-by-side analysis of two or more concepts or entities along shared dimensions. Comparison pages exist to answer contrastive questions: “How does X differ from Y?” A well-formed comparison page links to every subject it compares and uses parallel structure.Examples: rag-vs-llmwiki, transformer-vs-rnn, bm25-vs-dense-retrieval

overview

A map page that connects several related concepts in a domain. Overview pages are often seeded by schema.json rather than extracted directly from a source. They provide entry points into a topic cluster and link outward to the concept and entity pages that belong to that domain.Examples: retrieval-augmented-generation-overview, attention-mechanisms-overview
Page kinds are enforced by the optional .llmwiki/schema.json file. Projects without a schema file treat every page as kind: concept by default. Run llmwiki schema init to create a starter schema, or llmwiki schema show to inspect the resolved policy for your project.

​
Compiled Page Frontmatter

Every compiled page carries YAML frontmatter. Here is a representative example showing all standard fields:

​
Epistemic Metadata

confidence

A number between 0 and 1 representing the LLM’s reported confidence in the synthesized page. Higher values indicate the page is well-supported by its sources with minimal ambiguity. Lower values - by default, below 0.5 - cause llmwiki lint to flag the page with a low-confidence warning and, if the review policy is active, hold the page for review rather than writing it to wiki/.

provenanceState

A categorical label describing how the page’s content was produced:
  • extracted - derived directly from one source
  • merged - synthesized from multiple sources
  • inferred - the LLM drew a conclusion not stated explicitly in any source
  • ambiguous - sources gave conflicting signals; the page reflects the compiler’s best synthesis
When multiple sources merge into one slug, the reconciled state is always merged.

​
The contradictedBy Field

contradictedBy is a list of page slugs whose content is in tension with this page. When the LLM detects that two pages make conflicting claims about the same subject, it populates contradictedBy on one or both pages.
A non-empty contradictedBy triggers a contradicted-page lint warning and, if enabled in the review policy, holds the page as a candidate for human review.

​
How Metadata Affects the Review Policy

The review policy in .llmwiki/config.json uses epistemic metadata to decide whether a compiled page is written directly to wiki/ or held in .llmwiki/candidates/ for review. Any enabled hold mode that fires on a page’s metadata causes it to be held:
Pages with confidence below lowConfidenceThreshold (default 0.5) are held. Pages with no confidence field are also held by default - set "treatMissingConfidenceAs": "ok" in config.json to let them pass.
The review policy is fail-closed. An unknown hold mode or a corrupt config.json aborts the compile rather than silently disabling review. If you edit config.json manually, verify it with llmwiki schema show before the next compile.
When multiple sources merge into one slug, epistemic metadata is reconciled: the confidence takes the minimum across all contributing sources, provenanceState is set to merged, and contradictedBy is the deduplicated union of all contributing sources’ contradiction lists. For more detail on the schema layer and the review policy configuration format, see Schema and Review Policy.