Everyone bolts a vector database onto their AI agent — chop your docs into fragments, embed them, and hope similarity search re-finds the right ones on every question. Google Cloud just published a simpler option: the Open Knowledge Format. Write your knowledge down ONCE, as a folder of linked markdown files the agent reads directly. This kit is the whole thing on five pages.

Get the free OKF Starter Kit

Your AI agent doesn't actually know your business — so most teams bolt on RAG: chop every document into thousands of fragments, turn each into an embedding, store them in a vector database, and at query time paste back whatever scores highest. It works, but the agent re-discovers your company from shredded paper on every question, and you pay for an embedding model, a vector DB, chunk-size tuning, and constant re-indexing. On June 12, 2026 Google Cloud published a simpler option for the knowledge that's actually STRUCTURED — your tables, metrics, runbooks, APIs: the Open Knowledge Format (OKF), an open, vendor-neutral spec (github.com/GoogleCloudPlatform/knowledge-catalog, v0.1). An OKF 'bundle' is just a directory of markdown files — one file per concept — with YAML frontmatter. The entire format: exactly ONE required field, type; recommended fields title, description, resource, tags, timestamp; and two reserved filenames, index.md (the front door / progressive disclosure) and log.md (change history). Concepts link to each other with ordinary markdown links, so the agent TRAVERSES your knowledge deliberately — start at a metric, follow it to the table it's computed from, to the dataset, to the runbook — instead of guessing from fragments. Retrieval guesses; traversal navigates. This free kit makes it concrete: (1) the bundle layout + a real index.md; (2) two copy-paste concept files (a Metric and the BigQuery Table it links to) you can edit in minutes; (3) the one-page spec cheat-sheet — required vs recommended fields, the reserved filenames, and an honest OKF-vs-RAG 'when to use which' (OKF for the structured backbone, RAG for the messy long tail, both together for the best agents — OKF is not a replacement, it's relief); and (4) the exact prompt that makes Claude Code / Cursor / Gemini CLI build an OKF bundle from your codebase, plus a Monday checklist. The takeaway: stop forcing structured knowledge through fuzzy search — write it down once, and let the agent walk it. (OKF is v0.1 Draft — expect the spec to evolve.)

Subscribe to Hyperautomation AI ReportGet the PDF freeKeyword: OKF

By submitting, you'll be subscribed to the Hyperautomation AI Report newsletter — one practical AI email per week. Unsubscribe anytime.

Free. No spam. Unsubscribe anytime.