The okf/ directory of GoogleCloudPlatform/knowledge-catalog contains the spec, a reference agent, sample recipes, and example bundles (GA4, Stack Overflow, Bitcoin).

The reference agent

Produces an OKF bundle in two passes:

• BQ pass — one concept per object from BigQuery metadata alone.
• Web pass — the LLM acts as a crawler: it fetches seed URLs via a fetch_url tool and follows outbound links that look like authoritative docs, then either enriches an existing concept, mints a references/<slug> concept, or skips. Safety rails: --web-max-pages cap and a same-domain --web-allowed-host filter; --no-web skips crawling.

enrich

python -m reference_agent enrich \
    --source bq --dataset <project>.<dataset> \
    --web-seed-file <seeds.txt> --out ./bundles/<name>

--concept <type>/<name> (repeatable) iterates on a single concept.

visualize

python -m reference_agent visualize --bundle ./bundles/<name>

Writes a self-contained interactive viz.html (Cytoscape.js graph + marked for markdown, both from CDN, no backend, no data leaves the page): force-directed graph colored by type, edges from cross-links, a detail panel with frontmatter + rendered body, a "Cited by" backlinks list, plus search, type filter, and switchable layouts.

Relevance to this project

The reference agent shows the producer and consumer ends of OKF working end-to-end. It is domain-coupled (BigQuery + web crawl). Our contribution is a portable producer/maintainer: the generic ingest / query / lint skills, usable in any project, over the same format. The visualize command is directly reusable on any conformant bundle — including this one.

Citations

1. OKF README — https://github.com/GoogleCloudPlatform/knowledge-catalog/tree/main/okf