Skip to content
Knowledge Trust Ladder

Context Workspace integrations

Below, the toolkit is lokf at 0.8.0, the floor the sidecar pins. The toolkit writes a bundle out in three forms:

  • Markdown files;
  • RDF, as Turtle, N-Triples, JSON-LD or TriG from lokf convert, and as the four files lokf export writes;
  • tables, as CSV, Parquet or SQLite from lokf tables.

A domain’s workspace adds graph.nq, one named graph per bundle, and a registry keyed by base_iri.

Find your case

Section titled “Find your case”
Role in the workspace What KTL does today Where a product fits
Artifact repository The sidecar’s release workflow attaches the bundle zip, its checksum and an attestation to a GitHub release Any repository that stores a file and sends a webhook on arrival: a generic repository in JFrog Artifactory, Sonatype Nexus or GitLab, or an OCI registry through ORAS. JFrog documents Artifactory’s webhook starting a GitHub Actions workflow through workflow_dispatch
Store knowledge-federate.py serve answers SPARQL over the workspace, and lokf serve over one bundle, in an in-process Oxigraph store A store that loads N-Quads and answers SPARQL 1.1 (below)
Agent surface knowledge-federate.py mcp is the lokf-registry MCP server over the workspace, lokf mcp serves one bundle, and the docent runs in any agent that installs it A store’s own MCP server or natural-language layer (below)
Catalogue or semantic layer lokf tables writes the concepts and their relations as tables A catalogue that takes tables or YAML rather than RDF (below)
Stable identifiers base_iri in each bundle’s index.md w3id.org, a permanent redirection service, for a base_iri that outlives its host, and a static host such as GitHub Pages for the files lokf export writes

RDF stores with SPARQL

Section titled “RDF stores with SPARQL”

The workspace’s graph is graph.nq, in N-Quads, and every SPARQL 1.1 store surveyed loads N-Quads with its named graphs: among them Oxigraph, the toolkit’s own engine, Apache Jena Fuseki, GraphDB, Stardog, OpenLink Virtuoso and Amazon Neptune. What separates them is hosting, scale and price, which their own documentation covers. Managed SPARQL exists on AWS, through Amazon Neptune and marketplace offerings such as GraphDB’s, and natively on neither Azure nor Google Cloud, whose graph services speak Gremlin or GQL. On those two clouds, a store runs on a virtual machine or a marketplace image, or the bundle is loaded as tables instead. Among the property-graph stores surveyed, only Neo4j’s neosemantics imports N-Quads with their named graphs, and only on self-hosted Neo4j, not on Aura. The others load the bundle as tables.

Catalogues and semantic layers

Section titled “Catalogues and semantic layers”

lokf tables projects a bundle to “one table per concept type (the nodes) plus a single relations table (the typed edges: source, predicate, target)”, as CSV, Parquet or SQLite. For BigQuery or Athena it writes the Parquet to the location given and prints the external-table statements that register it. Every catalogue below that takes tables can read that output. The tables carry the structure and the frontmatter. The bound predicates, the meaning a graph carries, stay in the RDF.

Product What a bundle can feed it What comes out
data.world (ServiceNow), whose platform SPARQL “is also the query language upon which the data.world platform is based” RDF upload takes Turtle, N-Triples and RDF/XML; TriG, N-Quads and JSON-LD are listed as not supported. Each dataset is a named graph SPARQL, and a hosted MCP server
Microsoft Fabric IQ and Purview the schema as “TTL, RDF/RDFS, or OWL”, with the caveat that they “don’t provide complete or lossless support for the full RDF or OWL specification”; Purview’s Data Map speaks the Apache Atlas 2.2 APIs Work IQ and Foundry IQ for Copilot; no SPARQL
Databricks Unity Catalog lokf tables output as tables; metric views in SQL or YAML Genie, and a genie_one_mcp server
Snowflake tables, and YAML for semantic views Cortex Agents, and a managed MCP server
dbt Semantic Layer YAML metrics over tables “AI tools like Claude and ChatGPT can connect to the Semantic Layer through the dbt MCP server”
Google Cloud Knowledge Catalog (formerly Dataplex Universal Catalog) the OKF bundle itself: Google’s announcement says Knowledge Catalog ingests OKF, and the OKF repository carries a connector that pushes one bundle per entry group the catalogue’s search and context APIs
TopQuadrant TQ Data Foundation, “Built entirely on open knowledge-graph standards” of RDF, OWL and SHACL RDF, in principle; its import pages could not be read SPARQL and MCP, per its documentation index

The OKF repository’s README names the places an OKF bundle can be consumed by reading the files: BigQuery, Dataplex, Vertex AI, Gemini, the Agent Development Kit, LangChain, Unity Catalog, Collibra, Obsidian, Notion and the static site generators.

Agent-facing retrieval

Section titled “Agent-facing retrieval”

A workspace’s agent surface is the sidecar’s lokf-registry MCP server, or a store’s own. GraphDB and Stardog, among others, bring an MCP server and a natural-language layer of their own, while Amazon Neptune’s MCP server speaks openCypher and Gremlin, not SPARQL. Microsoft 365 Copilot takes the docent as a skill (m365.md). Graph-building retrieval frameworks, such as Microsoft GraphRAG and Amazon Bedrock Knowledge Bases GraphRAG, extract a graph of their own from text and consume no bundle’s graph. They can still read the bundle’s Markdown as documents, as OKF expects.

Sources

Section titled “Sources”