What is OKF?

Open Knowledge Format (OKF) is an open specification for storing an organization's "knowledge" as a directory of Markdown files with YAML frontmatter, so that both humans and AI agents can write, read, exchange, and use that knowledge — without needing any SDK, database, or specialized tooling.

If you can cat a file, you can read OKF · If you can git clone a repo, you can share OKF

Background

OKF v0.1 was released on 12 June 2026 by the Data Cloud team at Google Cloud (Sam McVeety and Amir Hormati) as a way of turning the "LLM-wiki pattern" proposed by Andrej Karpathy into a portable, vendor-neutral standard.

The core idea of LLM-wiki is: instead of having an LLM re-discover raw documents from scratch every time a question is asked, let AI progressively synthesize knowledge into organized, pre-linked Markdown pages, then load only the relevant pages directly into context.

Core Components (see Part 3 for details)

Term	Short definition
Bundle	A directory of all knowledge files — the unit of distribution
Concept	One unit of knowledge = one `.md` file (e.g. a table, a metric, a playbook)
Concept ID	The file's path within the bundle with `.md` stripped, e.g. `tables/orders.md` → `tables/orders`
Frontmatter	The YAML block at the top of a file (stores metadata such as `type`, `title`, `tags`)
Link	A Markdown link between concepts = builds relationships into a graph

Example: a single concept file

File tables/orders.md:

---
type: BigQuery Table
title: Orders
description: One row per customer order
tags: [sales, orders]
timestamp: 2026-06-15T00:00:00Z
---

# Schema
| Column | Type | Description |
| :--- | :--- | :--- |
| order_id | STRING | Order identifier (unique) |
| customer_id | STRING | FK to [customers](customers.md) |

# Joins
Joined to [customers](customers.md) via `customer_id`

As you can see, this is plain Markdown that is immediately human-readable — just a small YAML header and links to other concepts.

3 Design Principles

Minimally opinionated — frontmatter requires only a single type field; everything else is defined by the producer.
Producer and consumer are independently decoupled — a bundle written by hand, generated by an agent, or exported by a pipeline can be read by any tool.
A graph, not just a tree — concepts link to each other via Markdown links, creating richer relationships than folder structure alone.

What OKF is NOT

Not a fixed taxonomy — it does not prescribe which types must exist.
Not an opinion on how knowledge must be stored, served, or searched.
Not a replacement for domain-specific schemas (Avro, Protobuf, OpenAPI) — OKF references those; it does not absorb them.

Next, let's look at why you would choose OKF over traditional RAG → Why OKF?