The biggest obstacle to getting large language models to work inside a company isn’t the model itself — it’s the lack of internal background knowledge. Things like table schemas, workflow diagrams, and API documentation are scattered across different systems, buried in code comments, or locked in someone’s head. That means developers end up building custom retrieval pipelines for every new AI tool.
Enter Google’s Open Knowledge Format (OKF) v0.1, which formalizes the “LLM-Wiki” pattern first proposed by AI researcher Andrej Karpathy. The spec doesn’t require new software or SDKs — it simply asks teams to write their internal knowledge as plain Markdown files and store them in the code repository alongside the source.
Under the hood, an OKF knowledge pack is just a directory tree of Markdown files. Each file starts with a frontmatter-like block using text tags to declare its type, description, and date. Cross-references between files are handled by good old hyperlinks. A top-level index.md gives a quick overview of the architecture, and a log.md tracks update history.
Karpathy has noted that humans hate updating wikis — maintaining a table of contents and keeping cross-file links in sync is mind-numbingly tedious. But an AI never forgets to update a cross-reference and can edit up to 15 files at once. That makes it the ideal “document manager” for everyday wiki upkeep.
The spec also includes a forgiving parsing model. If the AI botches a field, types the wrong tag, or leaves a broken link, the system doesn’t crash. It just keeps going — a deliberate design choice for real-world reliability.