Author(s): Gaurav Shrivastav

Originally published on Towards AI.

Everyone saw the video model. Almost nobody read the 12-page spec that quietly reshapes how agents get their facts.

Google shipped two things in June that could not look more different.

The article argues that while Google’s Gemini video model gets attention, its Open Knowledge Format (OKF) is more important for how agents reliably obtain facts. It explains the problem OKF targets—fragmented internal knowledge spread across inconsistent sources and formats—contrasting typical RAG approaches with OKF’s “one canonical, human-readable, machine-parseable source of truth” stored as markdown files. Mechanically, OKF is described as a small set of rules: each concept is a file whose path is its identity, each file begins with YAML frontmatter (requiring only a type field), and concepts connect via ordinary relative markdown links to form a knowledge graph. The author then shows how OKF enables progressive disclosure: agents first read concise frontmatter “descriptions” for relevance, then load full bodies only when needed, avoiding token waste and making retrieval debuggable without embedding similarity scores. Next, the article demonstrates how to build a simple loader and “bundle” catalog, wire it into agent tools (list knowledge and fetch a concept by id), and solve the drift/fragmentation problem by keeping knowledge versioned like code. It positions OKF as complementary to MCP (which provides actions) and RAG (useful for large unstructured corpora), and warns about treating write access to the bundle like code to mitigate prompt-injection persistence. Finally, it briefly covers Google’s Gemini Omni Flash for conversational, editable text/image-to-video generation—showing how to iterate with a previous interaction id—then concludes that OKF is the quieter “contract” likely to outlast models because it governs how knowledge is represented and consumed.

Read the full blog for free on Medium.

Join thousands of data leaders on the AI newsletter. Join over 80,000 subscribers and keep up to date with the latest developments in AI. From research to projects and ideas. If you are building an AI startup, an AI-related product, or a service, we invite you to consider becoming a sponsor.

Published via Towards AI