We measured our AI knowledge base against Google's new format. Plain files won.
In June, Google published something called the Open Knowledge Format: a spec for how a company's AI knowledge base should be laid out. We found it in mid August, two months late, which stung enough that we now watch vendor publications on a schedule instead of trusting the feed to surface them.
Then we did the obvious thing. We ran our own AI assistant's knowledge base against the spec.
What OKF actually is
Almost nothing, in the best way. A directory of plain markdown files. One concept per file.
A small YAML header on each file that says what kind of knowledge it holds. Standard markdown links between files, which together form the knowledge graph. One index file listing what is in the corpus.
That is the spec. There is no product attached, no required tooling, no format you need a vendor to open. Every file is readable with cat, or Notepad, or any AI model ever trained.
Our score: 124 of 125
Our assistant has been building its knowledge base for months, file by file, with no spec in hand. When we measured it against OKF, 124 of 125 files were already the right shape. The gaps were one-line fixes: a header key sitting one level deeper than the spec wants it, and wiki-style links where standard markdown links belong.
We did not get lucky. Both designs grew from the same constraint. When the reader is a language model, the format that survives is the one every model already reads: plain text, light structure, human-legible.
We arrived there by trial and error. Google arrived there by publishing a spec. Same destination.
What this means if you are buying
Here is the part for business owners, because a standard like this shifts the ground under a sales pitch.
If a vendor offers your company an AI knowledge base that lives inside their product, in their format, ask one question: can I take my files and leave? If the answer involves an export tool, that is a no. Your company's operating knowledge is the asset. The tool reading it is the commodity.
Google publishing OKF makes the boring answer official: a company knowledge base should be a folder of markdown files you own, on storage you control, readable by whichever AI you point at it this year and whichever one you switch to next year. Grain stored in your own silo, not the mill's.
We sell knowledge repos built exactly this way, so yes, we are pleased the standard agrees with us. But the reasoning stood before the spec did, and you can verify it yourself with a text editor. That is the whole point.