Google cloud sent open knowledge format this week. It looks so simple: a directory with markdown files, each with a yaml frontmatter at the top, and a regular markdown link。
It's worth writing because it's painful. Agent has been able to write sql, check files, call tools, but is often stuck on the other level in the enterprise: whether or not this table is working, who ordered the calibration, how the order form and the customer's form should be connected, where the accident ranbook is, and whether a rule is sourced. The model does not lack grammar; what is lacking is the knowledge of the organization that can be consumed。
What is it
The basic unit of the okf definition is knowledge bundle. Bundle is a folder that can be placed in the git repository or can be called zip or tarball. Every non-reserved . Md file in the folder is known as concept and represents a knowledge unit。
CoNcept can be a bigquery table, a data set, an api endpoint, an indicator, a business process, an incident playbook, or a reference. File path is its id, such as the co of tables/orders. MdNcept id is tables/orders。
This design is quite engineering. You don't need to deploy services first, you don't need to register. If you can read the file, you can read the okf; if you can give the message, you can distribute the okf。
Technical details
One coNcept file divided into two paragraphs: yaml froNot matter and markdown body。
FroThe only mandatory field in ntmatter is type. It tells consumers this, coWhat about, like, bigquery table, metric, playbook, reference. Google doesn't centralize type registry, producers name themselves, consumers encounter unknown type should continue to read。
Recommended fields are title, desCription, resource, tags, timestamp. The idea here is clear:Ntmatter only contains a small amount of structured information that requires filtering, displaying, indexing。
The real semantics are in markdown body. For example, the description of the table's fields, examples of sqls, join methods, limitations on use, sources of reference can be written in headings, tables, code blocks, ordinary paragraphs. Spec gives several agreed titles: #schema, #examples, #citiations. It's part of the deal. It's good for people and agent。
Cross-file relations are also simple: regular markdown links. Orders. Md can be linked to /tables/customers. Md, the type of relationship is described in the surrounding text. This option reduces the cost of writing and allows graphic visualizers to treat all links as directed edge。
Here's the deal. Relationship semantics are not particularly strong, and it's too loose to do rigorous knowledge mapping. But i think it makes sense. When business knowledge was first organized, the hardest part was to make people want to write, to make it difficult to write. Strong schema is too early to press up, and many teams will just give up。
It's critical. Md and log. Md
Okf has two saved file names: index. Md and log. Md。
Index. Md for layer navigation. Agent can read the root directory in index, see what dataset, table, metric, playbook, and then open the file layer by task. There is no need to put the entire knowledge base in the context。
Log. Md is responsible for changing history. For example, a table has been added to a table of contents, a calibration of indicators has been modified, a playbook has been abandoned and can be recorded by date. After reading it, agent can see what has happened recently without having to rely entirely on full-text searches for luck。
The design clearly inherited the taste of llm wiki. Karpathy's llm wiki is also speaking on three floors: raw sources, wiki, schema. What okf does is to turn the model that has already emerged in this community into a supplier-neutral exchange format。
Why would google do this
Google's entry points are the data directory and the agency data club. There are too many contexts in the enterprise data system for service: dataplex, bigquery, loocker, documentation station, internal wiki, worksheet, slack discussion. Agent, if you really want to analyze data, it's not enough to get a field name。
A very common scene. A sales revenue indicator, a field name, may be called revenue, but it does not contain refunds, taxes, discounts, cross-border exchange rate adjustments, often in another document. Asked whether this indicator could be aggregated by week, whether it could be merged with the user table join, whether there were any data quality accidents in the historical division and spread to more places。
The traditional rag can retrieve this information, but it has to be respelled every time. Okf's idea is to organize stable knowledge into wikis, so it can accumulate. Agent's query is read by sorted concept, not just by raw docs。
Google also gave reference this time. Readme says that enrichment anent will read bigquery mEtadata, generate okf documents for concepts such as datasets, tables, etc.; then grab authoritative documents by web pass, and add to the status, description, relationship. It can also generate a self-incorporated HTML view of the co-in the bookleNcept and links。
What i care more about is that it pulls data catalogue back into the engineering stream. The indicator calibration update, the join path revision, and the addition of verified sql can all be git diff, go pr and be reviewed. This is much more practical for the data team than another chat portal。
How about a few adjacent criteria? Min
Okf can easily be seen with allms. Txt, mcp, openlineage, but they have different layers。
Llms. Txt is more like a site portal navigation for llm to tell where the model key documents are. Mcp is the protocol for agent connecting tools, resources and data sources. Openlineage records blood events while the operation is running, e. G., which job read which datasets and which ones produced。
Okf more like context asset itself. It can be directed by llms. Txt, it can be exposed to agent through mcp, and the openlineage result can be written into a concept statement or a definition. Its focus is on making organizational knowledge mobile, reviewable and readable document packages。
Landing
If the team wants to try the okf, i don't suggest that we take the entire company's knowledge base right up. A more stable path is the selection of a small area, such as a bigquery dataset, a core business indicator domain, or a data asset for a product line。
Local prefix. That's the type list, the recommendNtmatter fields, fixed chapter templates, civilisation rules. For example, tables must include schema, examples, citations; indicators must indicate caliber, filtering conditions, refreshing frequency, owner; playbook must specify trigger conditions, process steps, and route of upgrading。
The first version is then exported from the existing system. Dbt docs, lokml, bigquery information schema, datshub, dataplex, and internal wiki can all be producer. The first edition doesn't need to be perfectNcept made it。
And then you get people to review what's most critical. In particular, indicators calibration, permission boundaries, examples of sql, operational rules. Agent wrote something that could speed things up, but not as a source of fact. The status in the okf is important, and it is desirable that every key judgement can be traced to the original document, system link or owner。
After this step, get agent to read only consumption. For example, sql agent reads index. Md first, then opens the relevant tables and indicators concept, finds the join path along the links and the example query. See if the mass of the generation is stable and if errors are easier to locate. After reading only that consumption is stable, consider having agent write back the suggested changes and pass the pr for review。
Border
Okf v. 0. 1 is early. It has no access model, no search infrastructure, no uniform type registry, and no strong semantics. It is not a substitute for data governance systems, for schema management or for running blood collection。
Competencies should be especially careful. Dissemination becomes easier when table descriptions, examples of sqls, business rules are exported as documents. Which fields are written in the bookle, which examples leak business sensitive information, which coNcept can only read to a specific team, which is handled in the external system。
There is also the issue of scale. Hundreds to thousands of concepts, comfortable with files, git, index, grep, static viewer. At the hundreds of thousands level, there must be a search for indexes, fractions, incremental synchronization and permission filters. Okf is responsible for exchange patterns, not for large-scale servicing。
