Open Knowledge Format

An open specification from Google Cloud's Data Cloud team for keeping curated knowledge as a directory of markdown files with YAML frontmatter and ordinary links between them, so that people, AI agents and other tools can read the same files without a translation layer.

What is reusable: open sourceopen datano APIopen participation Apache 2.0A file format, not a service: no API, account or SDK is needed to read or write a bundle. Issues, pull requests and other implementations are invited.

Google Cloud's Data Cloud team announced the Open Knowledge Format (OKF) on 12 June 2026, in a post by Sam McVeety and Amir Hormati. They call it a formalisation of the "LLM wiki" pattern from Andrej Karpathy's gist of that name: a folder of markdown that a model or agent reads as context and keeps up to date. The makers say they wanted a format, not another service. OKF is tied to no cloud, database, model provider or agent framework.

A bundle is a directory tree. Each markdown file is one concept, such as a table, a dataset, a metric, a runbook, an API or an idea. The file's path is its identity. The frontmatter, a short YAML block at the top of the file, holds a few queryable fields. Only type is required. Recommended fields are title, description, resource (a URI for the asset the file describes) and tags.

Ordinary markdown links between files make the directory a graph. The surrounding prose, not the link, says what kind of relationship each link is. Two filenames are reserved: index.md, a listing of a directory for progressive disclosure, and log.md, a dated history of changes.

Version 0.2 of the spec adds optional families of fields:

provenance: a sources list with per-source signals such as author, usage count and last-modified date
trust: generated and verified, from which a reader derives a tier of unverified, machine-confirmed or human-reviewed
lifecycle: status and stale_after

A new concept type, Attested Computation, pins down how a figure is meant to be calculated. A consumer can then check a reported number against that recipe. Consumers must accept unknown types, unknown keys, missing optional fields and broken links.

Bundles ship as a git repository, an archive or a subdirectory of a larger repository. Reading or writing a bundle needs nothing beyond a text editor. Google published reference pieces with the spec:

an enrichment agent that reads a BigQuery dataset and writes one concept file per table or view, then on a second pass adds citations, schemas and join paths from documentation it crawls
a single-file HTML visualiser that shows a bundle as a graph
sample bundles for the GA4 e-commerce, Stack Overflow and Bitcoin public datasets

Google Cloud's Knowledge Catalog, formerly Dataplex, can ingest OKF bundles. The repository holding the spec and reference code is under the Apache 2.0 licence, with a disclaimer that it is not an official Google product.

The spec is at version 0.2. Its canonical home is now a repository of its own, GoogleCloudPlatform/open-knowledge-format; the copy under knowledge-catalog is a frozen snapshot. The makers call the format a starting point that will change as producers and consumers appear.

OKF keeps a vocabulary and its relations as plain files beside the data they describe, readable by a person and by a model. It fixes only the file conventions. Choosing the types and their relations is left to whoever writes the bundle, and is the work a method such as the Ontology Pipeline sets out.