Google Cloud ships Open Knowledge Format as a vendor-neutral spec for agent context

Google Cloud released the Open Knowledge Format (OKF) v0.1 on 2026-06-12, a vendor-neutral specification for representing the curated context that AI agents need at retrieval time. The format is deliberately small, markdown files with YAML frontmatter organized as a folder bundle, and the spec lives on GitHub under the GoogleCloudPlatform organization rather than being tied to a Google Cloud product. The framing in the launch post is that organizations stop rebuilding context plumbing per agent vendor and start producing knowledge artifacts that any agent harness can consume.

What shipped

OKF v0.1 is a directory of markdown files with YAML frontmatter following a fixed convention. Each file declares a type, title, description, and optional resource pointer in frontmatter; the markdown body holds the prose. Files cross-link by relative path, and folders nest as topical bundles such as datasets, tables, or concepts. The spec is a single page, the v0.1 release ships three sample bundles covering GA4 e-commerce, Stack Overflow, and Bitcoin datasets, and the entire format renders as ordinary markdown in any editor or on GitHub.

Two reference implementations land alongside the spec. The Knowledge Catalog enrichment agent walks a BigQuery dataset, generates OKF concept documents with citations and inferred schemas, and writes them into a target bundle. A static HTML visualizer turns an OKF bundle into an interactive graph view for exploration. Google Cloud’s existing Knowledge Catalog product updates to ingest OKF and serve it to downstream agents through the Catalog APIs, so OKF integrates with the Google Cloud surface without being gated by it. The repository accepts community implementations, and Andrej Karpathy’s articulation of the LLM-wiki pattern is cited in the launch post as the conceptual ancestor.

Where this lands in the market

The strategic pull is on the proprietary knowledge-graph and RAG-platform market. Every major vector database and agent platform ships its own ingestion format and a proprietary schema for representing curated knowledge, which means organizations standardizing on a vector store or an agent platform also lock in the way their knowledge artifacts are represented. OKF inverts that dependency by sitting upstream of the vector store: knowledge ships as portable markdown, and the embedding and retrieval layer becomes a swap rather than a content migration. The format’s deliberate simplicity (markdown, YAML, no SDK) is a direct response to Glean, Notion AI, and the enterprise RAG vendors who have been selling proprietary connectors as the differentiation.

For Google Cloud specifically, OKF is also a positioning move against Microsoft and AWS. Microsoft’s M365 Copilot and AWS Bedrock Knowledge Bases both treat the knowledge-ingestion layer as the lock-in surface for their respective ecosystems. Open-sourcing a portable format under a Google org rather than a neutral foundation has familiar tradeoffs; Google’s history with Angular, Flutter, and TensorFlow shows the pattern, but the format is small enough that a fork is trivial if Google’s stewardship goes sideways. The early Knowledge Catalog adoption signal matters because it is the first production agent system consuming OKF; the second-party adoption signal is where the spec earns its name.

What’s worth watching

1. Adoption by non-Google agent platforms. OKF only becomes a standard if LangChain, LlamaIndex, Mastra, Pinecone, Weaviate, and the agent frameworks built on them ingest OKF bundles natively. The early signal is whether a major agent platform adds OKF as an ingestion source in the next two quarters. Without that, OKF becomes another Google Cloud product format rather than the portable spec the launch promises.
2. Schema evolution and v0.2. The v0.1 spec is intentionally minimal, which is the right move for a launch but leaves real schema questions open: relationship typing between entries, embedding metadata, provenance, and versioning are all unaddressed. The shape of v0.2 will tell whether OKF stays a knowledge-as-markdown spec or grows toward the structured-data complexity of competing formats like RDF and JSON-LD.
3. Karpathy’s LLM-wiki concept maturing. The launch credits Karpathy’s LLM-wiki articulation as the conceptual ancestor. Whether Karpathy or other AI researchers publicly endorse OKF as the implementation of that pattern is the heat-signal for adoption beyond Google Cloud’s existing user base. A Karpathy tweet or an LLM-wiki repository moving to OKF format would matter more for industry positioning than the launch itself.

The plain reading is that Google Cloud is betting on a markdown-and-YAML spec to short-circuit the agent-platform lock-in race. The format is too small to fail technically; the question is whether it gets traction outside of the Google ecosystem in the next two quarters or stays a Knowledge Catalog ingestion convention. Both outcomes are plausible, and the difference depends less on the spec itself and more on whether one of LangChain, Mastra, or LlamaIndex publicly ships OKF support before Q4.

Sources cited

1. Google Cloud: Introducing the Open Knowledge Format cloud.google.com
2. Open Knowledge Format on GitHub github.com
3. Knowledge Catalog enrichment agent demo github.com