Open Knowledge Format: AI Knowledge as Markdown Files

Google Cloud has introduced Open Knowledge Format (OKF), an open specification designed to make context knowledge usable across platforms for AI systems and agents. The format is aimed at companies that want to centrally provide metadata, documentation, runbooks, or technical definitions for AI use. Google has already integrated support for OKF into its Knowledge Catalog.

With the format, Google is adopting an approach that has become widespread among AI agent developers in recent months. Instead of repeatedly having agents search for the same information from wikis, data catalogs, or documentation, teams store their knowledge in a structured way as a collection of Markdown files. AI researcher Andrej Karpathy has described this pattern as “LLM Wiki.” Related concepts are found in Obsidian vaults, in configuration files for agents like AGENTS.md or CLAUDE.md, and in so-called “Metadata as Code” repositories.

While these approaches use similar building blocks – Markdown files, metadata fields, and cross-references – they usually remain limited to individual teams, tools, or providers. This makes it difficult to reuse knowledge between different AI systems. This is precisely where Google comes in: OKF is intended to define the necessary conventions that allow different tools to read and write the same knowledge bases – without a translation layer and without proprietary vendor SDKs.

Knowledge as Markdown Files

An OKF bundle consists of a directory of Markdown files. Each file describes exactly one concept, such as a database table, a dataset, an API, a business metric, a runbook, or a playbook. The YAML frontmatter contains structured fields such as type, title, description, resource, tags, and timestamp.

The individual files link to each other via ordinary Markdown links, creating a knowledge graph that maps the relationships between concepts. For example, the documentation for an order table can reference customer and product data, as well as the definition of a revenue metric. This gives an AI agent not only individual documents but also their technical context.

Format instead of Platform

In the announcement, Google explicitly describes OKF as a format, not a platform. The specification is intended to function independently of cloud providers, databases, AI models, or agent frameworks. The developers have deliberately kept the standard lean: only a type field is mandatory; users can define all other structures and metadata themselves. Thus, OKF only dictates interoperability, not a uniform content model.

Along with the specification, Google provides several reference implementations. These include an enrichment agent for BigQuery that analyzes tables and views and automatically generates OKF documents from them. In a second pass, a language model enriches the documents with schema information, documentation, source references, and join relationships. Additionally, there is a static HTML viewer that displays an OKF repository as an interactive knowledge graph, entirely without a backend.

To try it out, Google provides example repositories for datasets from GA4 E-Commerce, Stack Overflow, and public Bitcoin datasets. The specification, example code, and reference implementations are available on GitHub.

The specification is currently in version 0.1. Google describes it as a starting point and plans to develop it further in a backward-compatible manner with the community. Producers and consumers of the format – such as data catalogs, search systems, or AI agents – are intended to emerge independently of each other.

(fo)