OKF - Simplified System Interaction (Part One)

Open Knowledge Format, an open format is a bit of a paradox, but it kinda works.

“Assumptions are the termites of relationships.”

-Henry Winkler

So, information, that’s important right? Knowledge is information recalled or put to action for a particular scenario. If we’re talking about LLMs, any LLMs, what people talk about is information, and how to get your information to the LLM in a way that represents the information correctly, has the model read it correctly and most importantly gives the correct result. This deterministic results from a non-deterministic system issue has existed and will eternally exist with AI.

OKF (Open Knowledge Format) is Google’s attempt to rectify that to an extent, and the lynchpin here is in its simplicity. It’s a format, meaning that it can be used with anything. Humans can read it and understand it, and AI can understand the strucuture and use it as a knowledge base with similar efficiency. Perhaps it is easier to exemplify:

---
type: Reference
resource: https://support.google.com/analytics/answer/9037342
title: Highly Active Users Metric
description: Builds an audience of users active for more than N minutes in the last M days.
tags:
  - metric
  - audience
  - ga4
  - high-actives
generated:
  by: reference_agent/gemini-3.5-flash
  at: 2026-07-10T21:16:29+00:00
sources:
  - id: sample_queries
    resource: https://support.google.com/analytics/answer/9037342
    title: Sample queries for audiences based on BigQuery data - Analytics Help
---

Builds an audience of Highly Active Users, defined as users who have been active/engaged for more than N minutes in the last M days, where M > N (for example, more than 0.1 minutes in the last 10 days).

# Schema
This reference describes a query pattern and does not map to a single database schema.

# Common query patterns

```sql
/**
 * Builds an audience of Highly Active Users.
 *
 * Highly Active Users = users who have been active for more than N minutes
 * in the last M days where M > N.
*/

SELECT
  COUNT(DISTINCT user_id) AS high_active_users_count
FROM
  (
    SELECT
      user_id,
      event_params.key,
      SUM(event_params.value.int_value)
    FROM
      `YOUR_TABLE.events_*` AS T
    CROSS JOIN
      T.event_params
    WHERE
      -- User engagement in the last M = 10 days.
      event_timestamp >
          UNIX_MICROS(TIMESTAMP_SUB(CURRENT_TIMESTAMP(), INTERVAL 10 DAY))
      AND event_params.key = 'engagement_time_msec'
      AND _TABLE_SUFFIX BETWEEN '20180521' AND '20240131'
    GROUP BY 1, 2
    HAVING
      -- Having engaged for more than N = 0.1 minutes.
      SUM(event_params.value.int_value) > 0.1 * 60 * 1000000
  );
```
[^sample_queries]

[^sample_queries]: [Google Analytics Help: Sample queries for audiences based on BigQuery data](https://support.google.com/analytics/answer/9037342)

You can find the full format that I have copied here and you may notice a couple of things: a) this was generated by Gemini and b) this is kind of like a blog, right? A blog that you could follow to make these queries if you were using BigQuery and support articles to help out if needed.

Here’s the beauty of that: good article writing might perhaps be critical to AI and Human interaction with technology in the future! Now, of course, much of these building blocks will be written using AI, but the space for creativity and order, which are two things which we don’t see as much in this area could result in both a new way of architecting infrastructure and a simplification of information and instructions fed to AI allowing less technical people the ability to provide input into AI design.

I’d recommend looking at the OKF article from Google which is a little outdated since we’re at v0.2 now, and if you’ve noted the part one on this blog, expect more on this soon!