llm-wiki-colab

Ontology, SHACL shapes, and SPARQL pipeline for llm-wiki-colab: a multi-agent, federated extension of the llm-wiki memory pattern with a typed-edge vocabulary and contribution / reconciliation primitives.

Serves the IRI namespace at https://la3d.github.io/llm-wiki-colab/ and ships the wiki → JSON-LD → Turtle build with materialized inverses and Jena-based SPARQL queries.

Why a new namespace

The llm-wiki pattern (Tobi Lütke / Shopify) is widely associated with that name. This project extends it with multi-agent attribution, cross-project contribution / reconciliation, and a richer typed-edge vocabulary, so it deserves a distinct prefix and IRI base to avoid collision.

Prefix: llm-wiki-colab:
IRI base: https://la3d.github.io/llm-wiki-colab/
- Ontology terms: <.../ontology#>
- Page IRIs: <.../page/{Page-Name}>

The IRI base is a GitHub Pages URL on a repo we control. Dereferenceable ontology documentation can be published there later; for now the IRIs are stable identifiers that resolve to a real repo.

Typed-edge vocabulary

Forward predicates the author asserts (in frontmatter or via inline body-link annotation) and the inverses the build pipeline materializes:

Forward	Inverse	Use
`up`	—	hierarchy parent (every page has one)
`source`	—	cites this external literature item
`extends`	—	builds on the intellectual content of the target
`supports`	`supportedBy`	provides evidence FOR the target’s claim
`criticizes`	`criticizedBy`	argues against the target’s claim
`concept`	`conceptOf`	discusses the target concept
`partOf`	`hasPart`	structural component or stage of the target
`dependsOn`	`prerequisiteOf`	design / argument depends on the target
`defines`	`definedBy`	canonical definition source for the target
`resolvedBy`	`resolves`	open question here is answered by the target
`incorporatedInto`	`incorporates`	content appears in the target (staging pattern)
`outOfScopeFor`	`excludes`	content explicitly NOT to appear in the target
`precedes`	`precededBy`	sequence order (version, stage, phase)
`feedsInto`	`informedBy`	signal / data flow into a pipeline
`related`	—	fallback when nothing more specific fits
`mentions`	—	body cross-reference, emitted automatically by the extractor

Inline body-link annotation form (Variant 1, agent-readable + self-documenting):

[Concept-X](Concept-X) ([*supports*](Edge-Types#supports))

The plain content link emits a mentions edge; the predicate link adds the typed edge. Predicate-carrier links to Edge-Types#* are excluded from mentions to keep that page from becoming a spurious hub.

Pipeline

wiki/*.md                     ← author-edited Markdown with YAML frontmatter
   │
   ▼   wiki-to-jsonld.py      ← extracts frontmatter + body-link annotations
*-graph.jsonld
   │
   ▼   riot (Apache Jena)     ← JSON-LD → Turtle
*-graph.ttl  +  *-weights.ttl ← weighted mentions as RDF-star
   │
   ▼   arq CONSTRUCT queries  ← materialize area inheritance + inverse edges + hubs
*-graph-full.ttl
   │
   ▼   shacl validate         ← *-shapes.ttl
validation-report.ttl

Built artifacts (*-graph*.ttl, *-graph.jsonld, validation-report.ttl) are generated, not checked in.

Quick start

Requires:

Python 3
Apache Jena CLI (riot, arq, shacl) — brew install jena

Build the KG from a wiki:

./build-graph.sh --wiki=/path/to/your.wiki --stats

Run a canned SPARQL query against the built graph:

arq --data=graph-full.ttl --query=sparql/hub-notes.rq

Layout

ontology.ttl       OWL/SKOS ontology (classes + predicates)
context.jsonld     JSON-LD context (prefix + edge term mappings)
shapes.ttl         SHACL shapes (frontmatter validation)
queries.ttl        Example queries in Turtle form (reference)
build-graph.sh     End-to-end pipeline
wiki-to-jsonld.py  Frontmatter + body-link extractor
sparql/            18 canned SPARQL queries (.rq)

File names are unqualified (no llm-wiki-colab- prefix) because the repo itself carries the namespace: when these files are dereferenced via GitHub Pages they land at https://la3d.github.io/llm-wiki-colab/ontology.ttl etc., matching the IRI base structure.

Attribution

The core ontology, JSON-LD context, SHACL shapes, and SPARQL pipeline are adapted from the working agentic-vault system developed by Dr. Charles Vardeman II (Research Assistant Professor, Center for Research Computing, University of Notre Dame). The base typed-edge vocabulary (up, source, extends, supports, criticizes, concept, related, plus associated classes and SHACL shapes) and the wiki → JSON-LD → Turtle → materialized inverses → SHACL pipeline shape are all his.

This repo’s additions, layered on that base for the multi-agent / federated case:

Typed-edge predicates: partOf / hasPart, dependsOn / prerequisiteOf, defines / definedBy, resolvedBy / resolves, incorporatedInto / incorporates, outOfScopeFor / excludes, precedes / precededBy, feedsInto / informedBy.
Inline body-link annotation form (Variant 1, visible predicate link to the Edge-Types page) so per-mention typed edges can sit next to plain prose links.
Contribution-status-aware verifier rules (canonical strict, contribution relaxed) for the federated wiki use case.
Namespace and IRI base for the collaborative variant (llm-wiki-colab: at https://la3d.github.io/llm-wiki-colab/).

License

MIT, see LICENSE.

This site is open source. Improve this page.