Open Knowledge Format (OKF): What It Is and How It Works

A new Domo content series is here, built for anyone who wants to move past basic reporting and into deeper platform mastery. The series covers hands-on tutorials, explores common challenges data teams face, and invites community participation to shape future topics. Whether you build pipelines, design apps, or translate insights for stakeholders, this program meets you where you are.

Key takeaways

• Definition: Open Knowledge Format (OKF) is an open specification that packages organizational knowledge as linked Markdown files with YAML (a human-readable metadata format) frontmatter, designed for AI agents to consume.
• Portability: The format lets you share governed knowledge across different AI tools without locking into any single vendor or platform.
• Position in the stack: OKF sits between your raw documentation and AI retrieval systems, acting as a structured context source for retrieval-augmented generation (RAG) pipelines and agents.
• Scope boundary: OKF is a format specification, not a hosting service or database. It standardizes structure, not storage.
• Maturity: The v0.1 spec is functional but early. Adoption patterns are still forming, which makes this a good time to experiment.

What is Open Knowledge Format (OKF)?

Google Cloud released Open Knowledge Format in 2024 to solve a problem that had been nagging AI teams for years: Agents need structured context to work accurately, but no standard existed for packaging and sharing organizational knowledge across tools. The format structures knowledge as a directory of interlinked Markdown files, each containing YAML frontmatter with metadata like titles, descriptions, and relationships.

Think of an OKF bundle as a folder. Inside, you have Markdown files where each file represents one concept (a metric definition, a policy, a product feature). A manifest file lists everything in the bundle. Each file's frontmatter declares how it connects to other concepts, creating a knowledge graph that AI systems can traverse.

What OKF is not: a replacement for JSON-LD, Schema.org, or semantic layers. Those serve different purposes. Web markup for search engines, metric definitions for BI tools. OKF targets AI agent context specifically. It's also not a wiki product. It's a format that wikis could export to.

The format has a few defining characteristics:

• Vendor-neutral: Bundles work with any system that reads the format.
• Human-readable: Markdown stays accessible to people, not just machines.
• Linked structure: Concepts reference each other, forming a traversable graph.
• Metadata-rich: YAML frontmatter carries provenance, versioning, and relationship data.

Why Open Knowledge Format matters for AI context

An AI agent gets asked to explain your company's revenue recognition policy. It hallucinates an answer because it can't access the finance team's documentation. According to the Stanford HAI 2026 AI Index Report, hallucination rates range from 22 percent to 94 percent across the top models they surveyed. Even the best-performing models fabricate answers roughly one-fifth of the time when they lack proper context. Or worse, the agent pulls from three different internal sources with conflicting definitions and blends them into one confident-sounding response.

Your organizational knowledge lives everywhere. Wikis, Notion pages, Confluence spaces, Google Docs, internal databases. This kind of fragmentation explains why AI agents so often return inconsistent answers. Each system structures content differently.

When teams build RAG pipelines or agent workflows, they spend enormous effort normalizing this fragmented content. They often lose metadata like authorship, update dates, and concept relationships in the process.

OKF addresses this by giving knowledge producers (documentation teams, subject matter experts) a standard structure. They package content once. Knowledge consumers (AI agents, search systems) ingest it without custom parsing. The format carries provenance metadata, so agents can cite sources and surface when content was last updated.

Not every situation calls for OKF. If your knowledge already lives in a semantic layer or data catalog that your AI tools consume well, adding another format creates overhead without clear benefit. Same goes for high-frequency transactional data. OKF is built for curated, narrative knowledge, not streams that update by the second.

But if your AI agents pull from multiple internal sources? Or you need to share governed context with partners without exposing your full knowledge base? Or you want the ability to switch AI platforms without rebuilding context?

How Open Knowledge Format works

Data teams evaluating a new spec immediately want to know what they will actually build and maintain.

A bundle is a directory containing Markdown files. Each file represents one concept. The bundle includes a manifest file listing contents and optional reserved filenames for metadata. Each file starts with YAML frontmatter: title, description, related concepts, last updated date, author.

Concepts reference each other through standard Markdown links or frontmatter relationships. An AI agent traversing the bundle follows these links to gather related context, similar to how you might click through wiki links to understand a topic. Teams sometimes create circular references between concepts without realizing it, which can cause some retrieval systems to loop or return duplicate content.

Three design principles shape the spec: 1) It's minimally opinionated, defining structure but not content style or domain vocabulary. 2) It maintains producer/consumer independence, meaning teams creating bundles don't need to know which AI systems will consume them. 3) And it's a format, not a platform. No specific database or hosting vendor required.

The lifecycle follows a predictable path:

1. Author: Subject matter experts write or export content as Markdown with frontmatter.
2. Validate: Run the bundle through a validator to catch structural errors.
3. Publish: Host the bundle anywhere (Git repository, static file server, cloud storage bucket).
4. Consume: AI agents, search systems, or other tools ingest the bundle as context.

Validation tooling exists in the official repository, though the ecosystem is early.

{{custom-cta-1}}

Open Knowledge Format use cases for data and AI teams

A data team maintains metric definitions in Confluence. Their AI assistant keeps returning inconsistent answers because it's pulling from cached, outdated pages. Sound familiar?

Here is where standardized packaging changes outcomes:

• Metric definitions: Analytics teams export governed definitions as an OKF bundle. AI assistants query it for authoritative answers about what "active user" or "monthly recurring revenue" means.
• Data lineage documentation: Data engineering packages lineage docs as linked concepts. When an agent answers questions about data sources, it traces relationships and cites the bundle.
• Compliance and policy context: Legal teams export policy documents with clear versioning. AI systems ground responses in current versions and flag outdated content.
• Cross-team knowledge sharing: A product team publishes feature documentation as a bundle. Partners consume it without direct access to internal systems.
• Dataset catalog supplements: Data catalogs often lack narrative context. Teams supplement catalog metadata with OKF bundles containing usage guidance and known issues.

Different tools serve different purposes. Internal wikis offer familiar editing but no standard AI export format. Semantic layers handle BI metrics well but aren't designed for narrative knowledge. Data catalogs track technical metadata but often lack contextual documentation. OKF bundles provide portable, AI-ready structure, though they require authoring and maintenance effort.

Choose OKF when you need portable, governed narrative knowledge that multiple AI systems can consume without custom integration work.

How OKF affects AI agents and search

Agents increasingly need to cite sources, explain reasoning, and operate on governed context. Not just generate plausible text. OKF addresses this by packaging knowledge with provenance metadata that agents can surface.

When an agent receives a query, it retrieves context from available sources. If that context comes from an OKF bundle, the agent accesses not just content but metadata: When it was updated, who authored it, what related concepts exist.

Teams building RAG systems often struggle with context quality. OKF bundles provide pre-structured, validated content that reduces parsing and normalization work in the ingestion pipeline. That said, OKF doesn't guarantee retrieval accuracy on its own. Your chunking strategy, embedding model, and retrieval logic still determine whether the right content surfaces for a given query. Those implementation choices often determine whether the right content shows up for a given query.

OKF doesn't replace vector databases, embedding models, or retrieval infrastructure. It standardizes source content.

Get started with Open Knowledge Format

You don't need to overhaul your knowledge infrastructure. Start with a single bundle covering one domain (metric definitions, a product area, a policy set) and expand from there.

1. Review the official spec: Read the v0.1 specification to understand required fields and structure.
2. Examine sample bundles: The official repository includes examples showing how concepts link together.
3. Create a minimal bundle: Start with three to five related concepts to test the structure.
4. Validate before publishing: Use the reference validator to catch structural errors.
5. Choose a hosting approach: Git repositories, static file servers, or cloud storage all work.

Read the OKF v0.1 specification

The spec is concise but technical. Focus on required frontmatter fields (title, description), optional fields (related concepts, last updated), and reserved filenames. The spec distinguishes between normative requirements (must follow) and informative guidance (recommended patterns).

The v0.1 release is current, and the spec will evolve.

Create and publish your first bundle

Create a folder. Add three Markdown files with frontmatter. Add a manifest file listing the concepts. Simple.

Run the reference validator command-line interface (CLI) tool against your bundle directory. You will hit errors. Missing required fields, broken internal links, malformed YAML. Fix errors before publishing.

Publishing options include Git repositories (AI systems can clone or fetch), static hosting (any file server or CDN), and cloud storage (Amazon S3, Google Cloud Storage (GCS), or Azure Blob Storage with appropriate access controls).

Teams often skip validation during content updates, leading to broken bundles. Adding validation to CI/CD pipelines catches errors before they reach production. Teams often lose hours debugging retrieval issues that trace back to a single missing frontmatter field.

How Domo supports Open Knowledge Format workflows

OKF standardizes how knowledge is packaged. Teams still need to govern, maintain, and operationalize that knowledge before packaging it.

Domo's data catalog tracks dataset definitions, lineage, and ownership. This content can inform OKF bundles for AI consumption. Role-based permissions ensure exported knowledge reflects governed, approved content. Domo's AI agents operate with bounded autonomy on governed data, with human-in-the-loop controls and oversight. OKF provides a standard format for extending that governed context to external systems or partners. With connectors to documentation systems, databases, and cloud apps, Domo serves as a hub for aggregating knowledge before packaging.

OKF is an open standard. Domo does not lock teams into a proprietary format.

{{custom-cta-2}}

Final thoughts

OKF addresses a specific gap: AI systems need structured, portable knowledge, and no standard existed for packaging it. The format is early but functional. McKinsey reports only 23 percent of organizations are scaling agentic AI, which means teams experimenting now with standards like OKF will have a head start when more companies start using it.

Start small. Pick one area of knowledge, create a bundle, validate it, and see how your AI tools use it. The investment is low, and what you learn will be useful no matter how the spec changes..

AI agents increasingly need controlled information to work dependably. Standards like OKF help different AI systems work together. This reduces the need for custom setup, which often slows down how quickly companies adopt AI. Want to swap notes on what's working (and what's a little messy) as you test your first bundle? Join the Domo community and compare OKF experiments with other data and AI teams.