Introducing the Open Knowledge Format#

Original | Raw

Authors: Sam McVeety & Amir Hormati (Google Cloud, Data Analytics Engineering) Published: June 12, 2026

Summary#

Google Cloud introduces the Open Knowledge Format (OKF), an open specification that formalizes the llm-wiki-pattern into a portable, vendor-neutral, interoperable standard for representing metadata, context, and curated knowledge that AI agents need.

OKF v0.1 is deliberately minimal: a directory of markdown files with YAML frontmatter. The only required field is type. Everything else — what types exist, what fields to include, what sections the body has — is left to the producer.

Key Claims#

  1. The context problem is universal: Organizational knowledge is fragmented across catalogs, wikis, drives, code comments, and people’s heads. Every agent builder solves context-assembly from scratch.
  2. What’s missing is a format, not a service: The answer isn’t another knowledge platform — it’s an interoperability standard that anyone can produce/consume without an SDK.
  3. OKF formalizes existing patterns: Obsidian vaults, AGENTS.md/CLAUDE.md convention files, index.md + log.md artifacts, and “metadata as code” repos all use the same shape. OKF makes them interoperable.
  4. Explicitly cites andrej-karpathy’s LLM Wiki gist as the intellectual origin of the pattern.

Design: One Screen#

An OKF bundle is a directory of markdown files. Each file is a concept (table, metric, playbook, API, etc.). File path = identity.

---
type: BigQuery Table
title: Orders
description: One row per completed customer order.
resource: https://console.cloud.google.com/bigquery?p=acme&d=sales&t=orders
tags: [sales, revenue]
timestamp: 2026-05-28T14:30:00Z
---

Concepts link via standard markdown links → directory becomes a knowledge graph. Optional index.md for progressive disclosure; optional log.md for change history.

Three Design Principles#

  1. Minimally opinionated — Only type is required. The spec defines the interoperability surface, not the content model.
  2. Producer/consumer independence — Who writes and who reads are cleanly separated. Human-authored bundles consumed by agents; LLM-generated bundles browsed by humans; one LLM’s output queried by another.
  3. Format, not platform — No proprietary account, SDK, or cloud required. Value comes from how many parties speak it, not who owns it.

Reference Implementations#

  • Enrichment agent: Walks BigQuery datasets → drafts OKF concept per table/view → second LLM pass enriches with citations, schemas, join paths.
  • Static HTML visualizer: Turns any bundle into interactive graph view — single file, no backend, no data leaves the page.
  • Three sample bundles: GA4 e-commerce, Stack Overflow, Bitcoin public datasets.
  • Google Cloud Knowledge Catalog: Updated to ingest OKF and serve to agents.

Connections to This Wiki#

  • This wiki’s schema (.kiro/steering/llm-wiki-schema.md) is essentially a pre-OKF implementation of the same pattern — markdown + YAML frontmatter + cross-links + index.md + log.md. OKF formalizes what we’ve been doing.
  • OKF’s type field maps to our frontmatter type: source | entity | concept | analysis.
  • OKF’s progressive disclosure via index.md matches our wiki/index.md as entry point.
  • OKF’s “format not platform” principle aligns with icm-folder-structure’s “filesystem as architecture” thesis.
  • The fragmented context landscape problem is the same problem context-management addresses at the agent level — OKF solves it at the organizational/interchange level.

What’s New vs. Karpathy’s Gist#

Karpathy described the pattern. OKF provides:

  • A minimal interoperability contract (required fields, reserved filenames)
  • Conformance criteria for tooling to validate against
  • Producer/consumer separation as explicit design goal
  • Organizational scope (team wikis, metadata catalogs) vs. personal scope

See Also#

Backward Notebooklm Notes Guide Openai Agents Sdk Forward