---
title: Which Facts Should Your AI Agent Never Guess? RAG, OKF And The Governed Core
description: ClairX Blog about Agent Knowledge Management - OKF vs RAG. Blog written with GCP as cloud however the concept applies to building agents irrespective of platform including AWS, Azure, Anthropic, OpenAI, and such.
image: https://www.clair-x.ai/hubfs/AgentKnowledge.png
---

[Skip to main content](https://www.clair-x.ai/insights/which-facts-should-your-ai-agent-never-guess-rag-okf-and-the-governed-core#main-content)

[![ClairX\_logo\_White](https://www.clair-x.ai/hubfs/ClairX_logo_White.svg) ![ClairX\_logo\_Original](https://www.clair-x.ai/hubfs/ClairX_logo_Original.svg) ![ClairX\_logo\_White](https://www.clair-x.ai/hubfs/ClairX_logo_White.svg) ![ClairX\_logo\_Original](https://www.clair-x.ai/hubfs/ClairX_logo_Original.svg)](https://www.clair-x.ai/)

- [Services](https://www.clair-x.ai/enterprise-data-ai-services)
- [Company](https://www.clair-x.ai/about-us)
- [Insights](https://www.clair-x.ai/insights)

Open main navigation

Close main navigation

- [Services](https://www.clair-x.ai/enterprise-data-ai-services)
- [Company](https://www.clair-x.ai/about-us)
- [Insights](https://www.clair-x.ai/insights)
- [Contact Us](https://www.clair-x.ai/contact-us)

[Contact Us](https://www.clair-x.ai/contact-us)

[Blog](https://www.clair-x.ai/insights)

Tags

- [AI Readiness](https://www.clair-x.ai/insights/tag/ai-readiness)
- [Agentic AI](https://www.clair-x.ai/insights/tag/agentic-ai)
- [Enterprise AI](https://www.clair-x.ai/insights/tag/enterprise-ai)
- [UK Tech](https://www.clair-x.ai/insights/tag/uk-tech)
- [CDO Insights](https://www.clair-x.ai/insights/tag/cdo-insights)
- [Public Sector](https://www.clair-x.ai/insights/tag/public-sector)
- [AILeadership](https://www.clair-x.ai/insights/tag/aileadership)

Search

- [AI Readiness](https://www.clair-x.ai/insights/tag/ai-readiness)
- [Agentic AI](https://www.clair-x.ai/insights/tag/agentic-ai)
- [Enterprise AI](https://www.clair-x.ai/insights/tag/enterprise-ai)
- [UK Tech](https://www.clair-x.ai/insights/tag/uk-tech)
- [CDO Insights](https://www.clair-x.ai/insights/tag/cdo-insights)
- [Public Sector](https://www.clair-x.ai/insights/tag/public-sector)
- [AILeadership](https://www.clair-x.ai/insights/tag/aileadership)

[AI Readiness](https://www.clair-x.ai/insights/tag/ai-readiness) [Agentic AI](https://www.clair-x.ai/insights/tag/agentic-ai) [Enterprise AI](https://www.clair-x.ai/insights/tag/enterprise-ai)

# Which Facts Should Your AI Agent Never Guess? RAG, OKF And The Governed Core

![ClairX AI Engineering Team](https://www.clair-x.ai/hubfs/clairx-ai-team-avatar-1.svg)

 by [ClairX AI Engineering Team](https://www.clair-x.ai/insights/author/clairx-ai-team)

 Table Of Contents

- [RAG vs OKF: two answers to the question every AI agent asks](https://www.clair-x.ai/insights/which-facts-should-your-ai-agent-never-guess-rag-okf-and-the-governed-core#rag-vs-okf-two-answers-to-the-question-every-ai-agent-asks)
- [What RAG does well, and where it slips](https://www.clair-x.ai/insights/which-facts-should-your-ai-agent-never-guess-rag-okf-and-the-governed-core#what-rag-does-well-and-where-it-slips)
- [What OKF adds: facts written down once, with an owner](https://www.clair-x.ai/insights/which-facts-should-your-ai-agent-never-guess-rag-okf-and-the-governed-core#what-okf-adds-facts-written-down-once-with-an-owner)
- [The difference](https://www.clair-x.ai/insights/which-facts-should-your-ai-agent-never-guess-rag-okf-and-the-governed-core#the-difference)
- [Where each one belongs](https://www.clair-x.ai/insights/which-facts-should-your-ai-agent-never-guess-rag-okf-and-the-governed-core#where-each-one-belongs)
- [Running both together](https://www.clair-x.ai/insights/which-facts-should-your-ai-agent-never-guess-rag-okf-and-the-governed-core#running-both-together)
- [The risks leadership should weigh](https://www.clair-x.ai/insights/which-facts-should-your-ai-agent-never-guess-rag-okf-and-the-governed-core#the-risks-leadership-should-weigh)
- [Five questions for your teams](https://www.clair-x.ai/insights/which-facts-should-your-ai-agent-never-guess-rag-okf-and-the-governed-core#five-questions-for-your-teams)
- [The data foundation decides the answer](https://www.clair-x.ai/insights/which-facts-should-your-ai-agent-never-guess-rag-okf-and-the-governed-core#the-data-foundation-decides-the-answer)

 Table Of Contents

- [RAG vs OKF: two answers to the question every AI agent asks](https://www.clair-x.ai/insights/which-facts-should-your-ai-agent-never-guess-rag-okf-and-the-governed-core#rag-vs-okf-two-answers-to-the-question-every-ai-agent-asks)
- [What RAG does well, and where it slips](https://www.clair-x.ai/insights/which-facts-should-your-ai-agent-never-guess-rag-okf-and-the-governed-core#what-rag-does-well-and-where-it-slips)
- [What OKF adds: facts written down once, with an owner](https://www.clair-x.ai/insights/which-facts-should-your-ai-agent-never-guess-rag-okf-and-the-governed-core#what-okf-adds-facts-written-down-once-with-an-owner)
- [The difference](https://www.clair-x.ai/insights/which-facts-should-your-ai-agent-never-guess-rag-okf-and-the-governed-core#the-difference)
- [Where each one belongs](https://www.clair-x.ai/insights/which-facts-should-your-ai-agent-never-guess-rag-okf-and-the-governed-core#where-each-one-belongs)
- [Running both together](https://www.clair-x.ai/insights/which-facts-should-your-ai-agent-never-guess-rag-okf-and-the-governed-core#running-both-together)
- [The risks leadership should weigh](https://www.clair-x.ai/insights/which-facts-should-your-ai-agent-never-guess-rag-okf-and-the-governed-core#the-risks-leadership-should-weigh)
- [Five questions for your teams](https://www.clair-x.ai/insights/which-facts-should-your-ai-agent-never-guess-rag-okf-and-the-governed-core#five-questions-for-your-teams)
- [The data foundation decides the answer](https://www.clair-x.ai/insights/which-facts-should-your-ai-agent-never-guess-rag-okf-and-the-governed-core#the-data-foundation-decides-the-answer)

![](https://www.clair-x.ai/hubfs/AgentKnowledge.png)

*A briefing for leaders deciding where their AI agents get the facts they act on, and how those facts stay trustworthy.*

*A briefing for leaders deciding where their AI agents get the facts they act on, and how those facts stay trustworthy.*

## RAG vs OKF: two answers to the question every AI agent asks

 

Every AI agent you deploy has to settle one question before it can answer yours: where do the facts come from? Two approaches now compete for that job. Retrieval-augmented generation, or RAG, searches your documents at the moment a question is asked. The Open Knowledge Format, or OKF, published by Google Cloud on 12 June 2026, writes your most important facts down once, in a governed format the agent reads directly. Most enterprises will run both. The decision that belongs with leadership is which knowledge goes where, and who owns it.

Picture a familiar scenario. A finance director asks the company's AI assistant for last quarter's churn rate. The answer arrives in seconds, well written and confident. It is also wrong, by a margin small enough that nobody questions it: the assistant used a definition of "active customer" from a two-year-old slide deck rather than the one the board approved in March. Nobody notices for a fortnight.

The assistant had plenty of data. What it lacked was a way to tell the approved version from the stale one, and that gap is what the RAG versus OKF conversation is really about. Some headlines called OKF a "RAG killer" within weeks of launch. The two work at different layers of the stack, and treating them as rivals risks your teams rebuilding the wrong thing.

## What RAG does well, and where it slips

 

RAG solves a scale problem. A language model was never trained on your contracts, tickets or policies, so RAG splits those documents into small passages and, at question time, hands the model the passages that look closest to the question. The model writes its answer from them.

For large, fast-changing material, such as millions of support tickets or a decade of email threads, it works well, because nobody could curate that by hand. Our team has built and run these pipelines on large programmes, and for that long tail of unstructured content we still reach for it first.

Its weak spots come from the same design. Four of them matter at the board level:

1. It rewards resemblance. The passage that looks most relevant can still be the wrong one, and on a precise figure, close is wrong.
2. It can split context. A definition and its exceptions may end up in different passages.
3. It fails quietly. A wrong answer arrives as fluently as a right one, with little signal by default unless your team has built source citations and monitoring around it.
4. It keeps old documents alive. A superseded policy stays as retrievable as the current one until someone builds the logic to retire it.
5. Which twenty facts would hurt most if our AI assistant got them wrong? That list is your starting point, usually KPI definitions, regulated thresholds and the schemas behind the main reports.
6. Who owns each of those facts today? If the answer is "nobody" or "it depends who you ask", that is a data governance gap, and closing it comes before any choice of technology.
7. Can we see when our AI gives a wrong answer? If the current system cannot show which source it used and whether that source was current, you are flying without instruments.
8. Where does our knowledge live? It belongs in your own version control, reviewed like code, with the serving layer kept swappable.
9. Would this satisfy an auditor? The EU AI Act's obligations for high-risk systems call for documentation and human oversight, and India's AI Governance Guidelines, non-binding today, ask organisations to explain how AI systems reach decisions. A verified, versioned source of truth is easier to defend than a search index.

RAG is the right tool for most enterprise content. Its blind spot is the small set of facts the business has already agreed and needs exactly right, every time. Our churn definition sits squarely in that set.

 

## What OKF adds: facts written down once, with an owner

 

*OKF addresses a trust problem.* Your organisation writes each important fact once, as a short file, and the agent reads that file directly instead of reconstructing the fact from search results. A "knowledge bundle" is a folder of these files, one per metric definition, business rule, table schema or policy, linked to one another and kept in version control alongside your code. Changes go through the same review and approval as any code change.

Google describes it as a format rather than a product: open, vendor-neutral, released under the Apache 2.0 licence, and readable by any tool. It formalises a pattern Andrej Karpathy described in April 2026, where an agent works from a maintained set of pages rather than re-deriving knowledge on every query.

The part that matters most to leadership arrived in version 0.2, published on 24 July 2026: a trust layer. Each fact can now record where it came from, who produced it, whether a person or a machine verified it, whether it is still current, and when it should expire. A new concept type, Attested Computation, links a reported figure to the approved query behind it.

That gives governance something concrete to hold: a churn definition reviewed in July and set to expire in October looks different from an unverified note on SharePoint, to your agent and to your auditor. RAG systems can carry similar labels, but each team invents its own. OKF makes them a shared standard.

Two caveats sit alongside that. OKF is young: Google launched it with three sample bundles, and as of late October we have found no other major vendor that has formally adopted it. And the trust fields are advisory. They record evidence about a fact, while access control and fact-checking stay with the people and systems around them.

 

## The difference

 

**RAG is a way of finding information. OKF is a way of writing it down.** One governs how an agent looks things up; the other governs what it finds when it gets there. So "RAG or OKF?" is usually the wrong question for a steering committee. The better one is: which of our facts deserve to be written down?

| **Question a leader would ask** | **RAG** | **OKF** |
| --- | --- | --- |
| **What problem does it solve?** | Scale: a huge, changing body of documents | Trust: a small set of agreed facts |
| **What knowledge suits it?** | Large, unstructured, hard to predict | Curated facts you can name in advance |
| **How does the agent get an answer?** | Searches and ranks passages | Reads the approved fact directly |
| **Who owns and checks it?** | Whatever each team builds | Standard fields for source, verifier and expiry |
| **What does it cost to run?** | Search infrastructure and processing on every query | Expert time to write and review; more processing as the bundle grows |
| **How does it fail?** | Quietly, with a plausible wrong answer | Visibly, with a fact flagged as expired or unverified |
| **How mature is it?** | Widely deployed, mature tooling | Published June 2026; early ecosystem |

 

Look twice at the failure row. For a regulated business, a failure you can see is often worth more than a small gain in accuracy.

 

## Where each one belongs

 

Four questions about the knowledge itself settle most decisions faster than any vendor briefing. How big is it? How often does it change? What does a wrong answer cost? How often is it reused?

OKF suits knowledge that is small, stable, reused constantly and expensive to get wrong. From our team's work on banking, payments and analytics programmes, that list is shorter than most leaders expect, and it is almost always where an agent's mistakes do the most damage. RAG suits the rest: large, unstructured and changing faster than anyone could curate.

| **Knowledge** | **Best fit** | **Why** |
| --- | --- | --- |
| **KPI and metric definitions** | OKF | Used in most analytics questions; must match the board pack |
| **Table schemas and data catalogue context** | OKF | Helps an analytics agent write the right query first time; Google's reference tooling generates these from BigQuery |
| **Regulated rules: credit thresholds, KYC steps, claims criteria** | OKF | A human-reviewed audit trail an examiner can follow |
| **Approved runbooks and procedures** | OKF | Versioned, reviewed like code, with a named owner |
| **Support tickets, chat logs and contract libraries** | RAG | Too large and varied to curate by hand |
| **Research and market intelligence** | RAG | Changes daily; curation cannot keep pace |
| **An enterprise assistant across finance, risk and operations** | Both | Agreed facts to OKF, everything else to RAG |

 

## Running both together

 

The pattern we expect most enterprises to settle on is a hybrid with a lopsided split: a small curated core in OKF, and the large majority of content in RAG. A router in front of the agent, either a simple classifier or the agent itself, decides which path each question takes.

Questions about an agreed concept, such as "what was Q3 churn?", go to the OKF bundle for the verified definition and the query behind it. Open-ended questions, such as "what are customers complaining about this month?", go to RAG. Plenty of questions need both: the agent takes the churn definition from OKF, then searches support tickets to explain the movement.

![RAg OKF Slide1](https://www.clair-x.ai/hs-fs/hubfs/RAg%20OKF%20Slide1.jpg?width=725&height=408&name=RAg%20OKF%20Slide1.jpg)

***Figure 1.*** *The router sends questions about agreed concepts to the OKF bundle and everything else to RAG; both feed one answer that can show where it came from.*

The two strengthen each other: a clean OKF bundle is also good input for a RAG pipeline, so writing knowledge down properly improves search too. Cost points the same way. OKF avoids search infrastructure for the facts it holds, but loading thousands of files strains the agent's working memory, even with index files to narrow the search. A small, well-kept core in OKF and the bulk in RAG makes sense on governance and on cost.

 

## The risks leadership should weigh

 

RAG's risks are about accuracy you cannot see: it retrieves what looks right, and leaves governance labels to each team, which makes cross-team audit hard. Its failures tend to surface in a board pack or a customer complaint rather than on a dashboard.

OKF's risks are about *discipline* and *maturity*:

- **Upkeep.** A bundle nobody owns becomes an unmaintained wiki with the authority of an official source. Expiry dates make decay visible; fixing it still takes an owner.
- **Over-curation.** Teams that try to put everything into OKF rebuild the cost of a full knowledge-management programme. Our test is reuse under pressure: if an agent reads a fact on nearly every task and a wrong answer causes real damage, it earns a place. Everything else stays in search.
- **An unproven standard.** Support so far comes mainly from Google's own Knowledge Catalog, and the format may or may not be widely adopted. Because a bundle is plain text in your own version control, the curation work keeps its value either way.
- **Commercial gravity.** The format is free, and Google updated its paid Knowledge Catalog on the same day to read and serve it. Keep the bundle in your own repository and treat any serving layer as replaceable.
- **Access and accuracy.** The specification has no document- or field-level permissions, and its validation checks structure rather than truth. A bundle needs the same access controls and review as the data it describes.

Each approach is one component of governance, and both need named owners, review cycles and monitoring.

 

## Five questions for your teams

 

Start small: one domain, its core facts in OKF, RAG for the rest, and a measured comparison. These questions show where you stand.

A pilot follows that order: list the critical facts, assign owners, write them as an OKF bundle with verification and expiry set, and route those questions through it, leaving RAG for everything else. Then compare accuracy and untraceable errors before and after. It takes weeks rather than quarters, and it gives your steering committee evidence to decide on.

## The data foundation decides the answer

 

This is a familiar conversation with a *modern lens*. Every enterprise has always had a few facts it must agree on and a great deal of material to search; AI agents now act on both, at speed, and need to know which is which. Our view is straightforward. RAG remains the right way to search the long tail. OKF is a promising way to govern the core, and its trust layer fills a gap RAG left open. Underneath both sits the real work: deciding which facts matter, who owns them and how they stay current. That is data governance, and it comes first.

We build on the platforms our clients already run, whether that is Databricks, Snowflake, AWS, Google Cloud or Microsoft, and the knowledge layer we help design stays in your repositories, owned by your teams. If you are weighing where OKF fits in your AI roadmap, or why your assistant keeps getting the same numbers wrong, we would be glad to talk it through.

\- Published on November 14th 2026.

 

If you would like to discuss RAG and OKF in detail or share how you are handling this, please do write us a note on [contact@clair-x.ai](mailto:contact@clair-x.ai)

Like what you've read? Share this insight

- [![Share on LinkedIn](https://www.clair-x.ai/hubfs/raw_assets/public/Clair-X/images/linkedin.svg)](https://www.linkedin.com/sharing/share-offsite/?url=https://www.clair-x.ai/insights/which-facts-should-your-ai-agent-never-guess-rag-okf-and-the-governed-core)
- [![Share on Twitter](https://www.clair-x.ai/hubfs/raw_assets/public/Clair-X/images/x.svg)](https://twitter.com/intent/tweet?url=https://www.clair-x.ai/insights/which-facts-should-your-ai-agent-never-guess-rag-okf-and-the-governed-core)

[![ClairX\_logo\_White](https://www.clair-x.ai/hubfs/ClairX_logo_White.svg)](https://www.clair-x.ai/)

 ClairX engineers the final mile of AI, transforming enterprise data into reliable decision engines with engineering rigor, built-in governance, and rapid delivery to unlock measurable operational and commercial value.

- - [Services](https://www.clair-x.ai/enterprise-data-ai-services)

- - [Company](https://www.clair-x.ai/about-us)

- - [Insights](https://www.clair-x.ai/insights)

- - [Contact Us](https://www.clair-x.ai/contact-us)

© 2026 Copyright ClairX. All Rights Reserved. [Privacy Policy](https://www.clair-x.ai/privacy-policy)

Follow us: [![linked-in](https://www.clair-x.ai/hubfs/linked-in.svg)](https://www.linkedin.com/company/clair-x/)

```json
{
  "@context" : "https://schema.org",
  "@type" : "BlogPosting",
  "author" : {
    "@type" : "Person",
    "name" : "ClairX AI Engineering Team",
    "url" : "https://www.clair-x.ai/insights/author/clairx-ai-team"
  },
  "dateModified" : "2026-11-14T10:28:25.358Z",
  "datePublished" : "2026-11-14T10:28:25.000Z",
  "headline" : "Which Facts Should Your AI Agent Never Guess? RAG, OKF And The Governed Core",
  "image" : [ "https://www.clair-x.ai/hubfs/AgentKnowledge.png" ],
  "mainEntityOfPage" : {
    "@id" : "https://www.clair-x.ai/insights/which-facts-should-your-ai-agent-never-guess-rag-okf-and-the-governed-core",
    "@type" : "WebPage"
  },
  "publisher" : {
    "@type" : "Organization",
    "logo" : {
      "@type" : "ImageObject",
      "url" : "https://www.clair-x.ai/hubfs/ClairX_logo_Original.svg"
    }
  }
}
```