In June 2026, Google launched the Open Data Format (OKF), an open specification for a way AI brokers organise and trade data. An OKF bundle is simply Markdown recordsdata, light-weight YAML metadata, and hyperlinks between ideas, but it challenges the belief that each AI software wants embeddings and vector databases.
As a result of the data base is obvious textual content, it may be version-controlled in Git and navigated by following hyperlinks slightly than retrieving disconnected chunks. On this article, we’ll discover how OKF works and when it beats a conventional retrieval pipeline.
Why Conventional RAG Has Limitations
Over the previous few years, Retrieval-Augmented Era (RAG) has turn out to be the usual method for offering exterior data to Giant Language Fashions. As a substitute of relying solely on the mannequin’s coaching information, RAG retrieves related info from exterior paperwork throughout inference. A typical pipeline appears one thing like this:
This method works remarkably nicely for looking out hundreds of thousands of paperwork. By evaluating the semantic which means of embeddings as an alternative of tangible key phrases, RAG permits an AI system to reply questions utilizing info that was by no means a part of the mannequin’s authentic coaching information.
Nonetheless, there is a vital trade-off. Earlier than a doc may be listed, it should first be divided into smaller chunks. Whereas chunking improves retrieval effectivity, it additionally breaks aside the unique construction of the doc. Relationships that had been naturally linked inside a single doc turn out to be distributed throughout a number of unbiased chunks.
Take into account the next hospital protocol.
# Affected person Admission CoverageSufferers arriving by the Emergency Division should full an preliminary triage earlier than admission.
## Admission Necessities
- Legitimate affected person identification
- Preliminary scientific evaluation accomplished
- Emergency circumstances obtain fast precedence## Recording and Mattress Allocation
Affected person info is recorded within the Digital Well being File (EHR) system earlier than a mattress is assigned.
Mattress allocation follows the Mattress Occupancy pointers maintained by the Operations group.
A typical RAG pipeline might break up this doc into a number of smaller chunks earlier than indexing.
1st Chunk
Sufferers arriving by the Emergency Division should full an preliminary triage.Admission Necessities:
- Legitimate affected person identification
- Preliminary scientific evaluation
- Emergency circumstances obtain fast precedence
2nd Chunk
Affected person info is recorded within the Digital Well being File (EHR) system.
third Chunk
Mattress allocation follows the Mattress Occupancy pointers maintained by the Operations group.
Visually, the method appears like this:
When a clinician asks,
“What’s the affected person admission course of?”
The vector database retrieves the chunks that appear most related, however the logical relationships between the admission coverage, emergency triage, the EHR system, and mattress allocation are misplaced. The mannequin has to reconstruct them on each question. This isn’t a flaw in RAG. It stays probably the greatest methods for looking out giant, unstructured collections like PDFs, analysis papers, help tickets, and historic information.
Curated organisational data is totally different. Insurance policies, procedures, APIs, and runbooks aren’t simply textual content, they’re interconnected ideas. Rebuilding these hyperlinks from fragmented chunks on each question provides pointless complexity, and that’s precisely the issue OKF was designed to unravel.
What’s the Open Data Format (OKF)?
The concepts behind OKF didn’t originate with Google. Earlier in 2026, Andrej Karpathy launched the idea of an LLM Wiki: as an alternative of repeatedly retrieving uncooked paperwork, an AI agent maintains a curated data base it could possibly repeatedly learn, replace, and enhance. His analogy caught on rapidly within the AI neighborhood:
Obsidian is the IDE. The LLM is the programmer. The wiki is the codebase.
The concept is easy. People present supply materials like documentation, insurance policies, schemas, and runbooks, and the agent organises it right into a structured wiki by writing summaries, connecting associated ideas, and sustaining hyperlinks. These relationships turn out to be a part of the data base as an alternative of being rediscovered on each question.
Google turned this neighborhood concept into an open specification. Quite than transport one other framework or SDK, it targeted on standardising the data itself. The result’s OKF, a light-weight format that shops data as odd Markdown recordsdata with minimal metadata and specific hyperlinks between ideas.
An OKF bundle is only a listing of Markdown paperwork, every representing one idea similar to a coverage, API, division, runbook, database desk, or metric, linked by normal Markdown hyperlinks. In contrast to a vector database that infers relationships by embedding similarity, OKF preserves them explicitly, so an agent follows hyperlinks slightly than guessing.
As a result of all the pieces is obvious textual content, it matches current developer workflows: version-controlled in Git, reviewed through pull requests, and searchable with normal instruments. Subsequent, we’ll construct a bundle from scratch to see how it’s organised.
Construction of an OKF Bundle
Now that we’ve understood the motivation behind OKF, let’s take a look at how an OKF bundle is definitely organised. At its core, an OKF bundle is solely a listing of Markdown recordsdata. Every Markdown file represents one idea, similar to a hospital coverage, division, process, system, or operational metric. Each idea accommodates light-weight metadata adopted by structured Markdown content material. Associated ideas are linked utilizing normal Markdown hyperlinks, permitting each people and AI brokers to navigate the data base naturally.
The specification itself is deliberately minimal. It defines just a few conventions and avoids implementing a inflexible listing construction. This offers organisations the pliability to organise data in a manner that most closely fits their area whereas nonetheless producing bundles that may be understood by any OKF-compatible agent.
A typical OKF bundle accommodates the next parts.
| Part | Goal |
| index.md | Serves as the first entry level into the data base. It gives an summary of the out there ideas and helps brokers navigate the bundle. |
| CHANGELOG.md (Optionally available) | Information adjustments made to the data base over time, making updates clear and traceable. |
| Idea Information (.md) | Every Markdown file represents a single idea similar to a coverage, process, API, division, metric, or system. |
| YAML Entrance Matter | Shops metadata together with the idea sort, title, description, tags, possession, and final up to date timestamp. |
| Markdown Hyperlinks | Explicitly join associated ideas, remodeling the data base right into a navigable graph as an alternative of remoted paperwork. |
A Typical OKF Folder Construction
Though the OKF specification doesn’t mandate a specific listing structure, following a constant folder hierarchy makes the data base considerably simpler to keep up and navigate. The identical organisational ideas apply whatever the area.
The next examples reveal how totally different organisations can construction their data whereas following the identical OKF conventions.
1: Hospital Data Base
2: Software program Engineering Data Base
3: Manufacturing Data Base
Though these examples belong to utterly totally different industries, the underlying organisation stays remarkably related. Each bundle begins with an index.md file that serves because the entry level, an non-compulsory CHANGELOG.md for monitoring revisions, and a set of directories that group associated ideas collectively.
This consistency is considered one of OKF’s greatest strengths. As soon as an AI agent understands how one OKF bundle is organised, it could possibly navigate one other bundle constructed utilizing the identical conventions with little or no further adaptation.
Constructing an OKF Bundle
Now that we’ve explored the general construction of an OKF bundle, let’s construct one from scratch.
For the rest of this text, we’ll use a fictional hospital known as CityCare Hospital. Think about we’re constructing an AI assistant that helps medical doctors, nurses, and hospital directors reply operational questions. The assistant ought to perceive admission insurance policies, emergency procedures, hospital departments, inside techniques, and operational metrics. As a substitute of storing this info inside a vector database, we’ll organise it as an OKF bundle.
We’ll start by creating the basis listing.
The index.md file acts because the entry level for each people and AI brokers.
# CityCare Hospital Data Base## Insurance policies
- [Patient Admission Policy](insurance policies/patient-admission.md)
- [Discharge Policy](insurance policies/discharge-policy.md)## Procedures
- [Emergency Triage](procedures/emergency-triage.md)
- [Blood Transfusion](procedures/blood-transfusion.md)## Programs
- [Electronic Health Record](techniques/ehr-system.md)
## Metrics
- [Bed Occupancy](metrics/bed-occupancy.md)
## Departments
- [Emergency Department](departments/emergency.md)
Quite than looking out the complete repository, an AI agent can first learn the index to grasp what ideas exist earlier than navigating to the related recordsdata. This easy design retains the bundle organised whereas lowering pointless context throughout retrieval.
Within the subsequent part, we’ll create particular person idea recordsdata and study how YAML metadata, Markdown content material, and hyperlinks work collectively to make the data base comprehensible for each people and AI brokers.
Creating an OKF Idea File
The constructing blocks of an OKF bundle are idea recordsdata. Every idea represents precisely one piece of data, similar to a coverage, process, division, system, metric, or API. Conserving ideas targeted makes them simpler to keep up whereas permitting AI brokers to retrieve solely the knowledge they want.
Each idea file consists of two components:
- YAML Entrance Matter, which shops metadata concerning the idea.
- Markdown Content material, which accommodates the precise data together with hyperlinks to associated ideas.
Let’s create an idea file for the hospital’s affected person admission coverage.
---
sort: coverage
title: Affected person Admission Coverage
description: Tips for admitting sufferers into CityCare Hospital
tags:
- admissions
- patient-care
up to date: 2026-06-15
---# Affected person Admission Coverage
Sufferers arriving by the Emergency Division should full an preliminary triage earlier than admission.
## Admission Necessities
- Legitimate affected person identification
- Preliminary scientific evaluation accomplished
- Emergency circumstances obtain fast precedence## Associated Ideas
- [Emergency Triage](../procedures/emergency-triage.md)
- [Electronic Health Record](../techniques/ehr-system.md)
Discover that the file accommodates far more than plain textual content. The YAML part describes what sort of data this file represents, whereas the Markdown physique explains the idea intimately. Most significantly, the idea hyperlinks to different associated ideas contained in the data base. These hyperlinks rework remoted paperwork into an interconnected data graph that an AI agent can navigate.
Though OKF solely requires the kind area, including further metadata makes the bundle simpler to organise and preserve.
| Subject | Description |
| sort | Identifies the kind of idea, similar to coverage, process, system, or metric. That is the one required area within the present specification. |
| title | Human-readable title of the idea. |
| description | Abstract describing the idea. |
| tags | Key phrases that assist organise associated ideas. |
| up to date | Signifies when the idea was final modified. |
For the reason that metadata is saved in YAML, each people and AI brokers can rapidly perceive what a doc represents earlier than studying its full content material.
One of many greatest variations between OKF and conventional doc storage is that ideas are explicitly linked utilizing Markdown hyperlinks.
For instance, the admission coverage references the emergency triage process and the Digital Well being File (EHR) system.
These relationships are deliberately created by the creator. The agent doesn’t should infer them by semantic similarity as a result of they exist already contained in the data base.
One other Instance: Hospital Metric
Idea recordsdata should not restricted to insurance policies. The identical construction can describe operational metrics, inside techniques, departments, APIs, or runbooks.
Beneath is an idea describing the hospital’s Mattress Occupancy Fee.
---
sort: metric
title: Mattress Occupancy Fee
description: Share of inpatient beds at the moment occupied
tags:
- operations
- hospital
up to date: 2026-06-15
---# Mattress Occupancy Fee
The Mattress Occupancy Fee measures the share of inpatient beds at the moment occupied.
## Method
Occupied Beds / Whole Out there Beds × 100## Knowledge Supply
Hospital Info System
## Proprietor
Operations Division
## Associated Ideas
- [Emergency Department](../departments/emergency.md)
- [Patient Admission Policy](../insurance policies/patient-admission.md)
As a result of each idea follows a constant construction, an AI agent can rapidly perceive what the metric represents, how it’s calculated, the place the info comes from, and which different ideas are associated to it.
How AI Brokers Traverse an OKF Bundle
As soon as the data base is organised into interconnected idea recordsdata, retrieval turns into a lot easier than conventional doc search.
As a substitute of looking out 1000’s of doc chunks, an AI agent follows a structured navigation course of.
- Learn the index.md file to grasp the general data base.
- Establish probably the most related idea based mostly on the person’s query.
- Open that idea file.
- Observe hyperlinks to associated ideas every time further context is required.
- Generate the ultimate response utilizing solely the related ideas.
The traversal course of may be visualised as follows.
In contrast to a RAG pipeline, the agent doesn’t start by looking out an embedding index. It begins from a curated entry level and progressively explores solely the ideas which can be related to the present activity.
This method preserves the relationships between ideas whereas conserving the quantity of context despatched to the language mannequin comparatively small.
Why This Works Nicely for AI Brokers
The most important benefit of OKF is that it permits builders to organise data in the identical manner people naturally give it some thought. A health care provider studying the hospital’s documentation doesn’t randomly soar between unrelated paragraphs. They start with a coverage, comply with references to procedures, seek the advice of the related techniques, after which arrive on the info they want. OKF permits AI brokers to comply with this identical workflow.
As a substitute of reconstructing relationships from fragmented doc chunks each time a query is requested, the agent navigates an specific data graph the place these relationships have already been outlined. This makes the retrieval course of extra deterministic, simpler to audit, and considerably easier to keep up.
The place RAG Nonetheless Excels
At this level, OKF seems like a really perfect answer for organising data. It preserves relationships between ideas, retains all the pieces version-controlled, and permits AI brokers to navigate curated documentation with out counting on semantic search.
Nonetheless, OKF has an essential limitation. Somebody has to curate each idea.
Each coverage, process, system, metric, and division have to be written, reviewed, and maintained. This works nicely for authoritative organisational data, however it turns into impractical when the data base grows to hundreds of thousands of paperwork.
Take into account a hospital that has collected years of operational information.
- Hundreds of thousands of Digital Well being File (EHR) entries
- Medical analysis papers
- Scientific notes
- Affected person suggestions
- Incident reviews
- Inside emails
- Assembly transcripts
- Tools upkeep logs
Organising each considered one of these paperwork into fastidiously curated OKF idea recordsdata would require an infinite quantity of handbook effort. Even when AI brokers assisted with the curation course of, a lot of this info adjustments repeatedly and is healthier suited to semantic search.
That is the place Retrieval-Augmented Era (RAG) continues to be the popular answer.
RAG Answer
Quite than requiring paperwork to be manually organised, RAG indexes giant collections of unstructured information utilizing embeddings. When a query is requested, the system retrieves probably the most semantically related paperwork and gives them as context to the language mannequin.
For instance, contemplate the next questions:
- Has anybody encountered this MRI scanner error earlier than?
- Discover earlier incident reviews involving delayed laboratory outcomes.
- Summarise discussions concerning the new EHR rollout.
- Seek for all assembly notes mentioning affected person switch delays.
These questions can’t be answered from a small curated data base. As a substitute, they require looking out by 1000’s and even hundreds of thousands of paperwork the place the reply might exist wherever. That is precisely the kind of drawback RAG was designed to unravel.
The strengths of every method turn out to be a lot clearer when seen facet by facet.
| Function | OKF | RAG |
| Greatest for | Curated organisational data | Giant collections of unstructured paperwork |
| Data Supply | Markdown idea recordsdata | Uncooked paperwork |
| Retrieval | Deterministic navigation | Semantic similarity search |
| Infrastructure | File system + Git | Embeddings + Vector Database |
| Model Management | Native Git help | Requires re-indexing after updates |
| Relationships | Specific hyperlinks between ideas | Inferred from retrieved chunks |
| Scalability | Reasonable | Wonderful |
| Explainability | Excessive | Reasonable |
Neither method is universally higher than the opposite. They merely clear up totally different issues. OKF gives construction and precision. RAG gives scale and suppleness. This naturally raises one other query.
Do we actually have to decide on one over the opposite?
Luckily, the reply is no.
Hybrid Data Structure: Combining OKF and RAG
In observe, the best AI techniques use each OKF and RAG collectively.
As a substitute of treating them as competing applied sciences, trendy agent architectures use each the place it performs greatest.
A easy manner to consider that is the 80/20 precept.
- The 20% of organisational data that’s important, steady, and incessantly referenced is saved as an OKF bundle.
- The remaining 80% of enormous, unstructured info stays inside a conventional RAG pipeline.
This creates a layered data structure.
The router determines which data supply is most applicable for the incoming question.
Questions requiring authoritative and deterministic solutions are routed to the OKF bundle.
Examples embrace:
- What’s the affected person admission coverage?
- How is the Mattress Occupancy Fee calculated?
- Which system shops affected person medical information?
- What’s the emergency blood transfusion process?
Every of those questions has a single authoritative reply maintained by the organisation.
Alternatively, exploratory questions are routed to the RAG pipeline.
For instance:
- Has anybody encountered this MRI scanner error earlier than?
- Discover earlier incident reviews involving delayed laboratory outcomes.
- Summarise discussions concerning the EHR migration undertaking.
- Search assembly notes discussing affected person discharge delays.
These questions require looking out giant collections of historic paperwork slightly than consulting curated data.
This hybrid structure permits every system to deal with its strengths.
| OKF Strengths | RAG Strengths |
| Deterministic retrieval | Semantic retrieval |
| Curated and authoritative data | Large doc collections |
| Specific relationships between ideas | Finds info utilizing semantic similarity |
| Model-controlled with Git | Repeatedly indexes new paperwork |
| Simple to audit and preserve | Extremely scalable |
Maybe the largest benefit of this structure is that the language mannequin doesn’t must know the place the knowledge comes from. The agent merely requests the data it wants.
The routing layer decides whether or not that data ought to come from the OKF bundle or the vector database. Frameworks similar to LangGraph, LangChain, or LlamaIndex make this routing easy by permitting builders to construct workflows that select the suitable retrieval technique based mostly on the person’s question. Consequently, AI brokers acquire the precision of curated data with out sacrificing the flexibility to look huge collections of unstructured info.
In different phrases, the longer term will not be OKF versus RAG. It’s OKF plus RAG, working collectively as complementary layers in a single data structure.
Conclusion
The Open Data Format gives a easy, clear method to organise data for AI brokers. By representing it as interconnected Markdown paperwork, OKF retains organisational data simple to grasp, preserve, and version-control in Git.
It fits curated info like insurance policies, runbooks, and API docs, with out changing RAG, which nonetheless excels at semantic search throughout giant, unstructured collections.
Used collectively, the 2 cowl way more floor than both alone. In the end, understanding the place every method matches is what lets builders construct brokers which can be correct, explainable, and production-ready.
Steadily Requested Questions
A. It’s an open specification Google launched in June 2026 for a way AI brokers organise and trade data. An OKF bundle is a listing of Markdown recordsdata with light-weight YAML metadata and specific hyperlinks between ideas, so data lives as plain textual content alongside your code slightly than as vectors in a database.
A. RAG splits paperwork into chunks, embeds them, and retrieves probably the most semantically related items at question time. OKF shops data as interconnected idea recordsdata and lets an agent navigate by following author-defined hyperlinks. RAG infers relationships; OKF retains them specific.
A. No. They clear up totally different issues. OKF is greatest for curated, authoritative data like insurance policies, runbooks, and API docs. RAG is greatest for looking out giant, unstructured collections similar to EHR entries, incident reviews, and assembly notes. The article recommends utilizing them collectively.
GenAI Intern @ Analytics Vidhya | Last Yr @ VIT Chennai
Obsessed with AI and machine studying, I am desirous to dive into roles as an AI/ML Engineer or Knowledge Scientist the place I could make an actual affect. With a knack for fast studying and a love for teamwork, I am excited to convey modern options and cutting-edge developments to the desk. My curiosity drives me to discover AI throughout numerous fields and take the initiative to delve into information engineering, making certain I keep forward and ship impactful initiatives.
Login to proceed studying and luxuriate in expert-curated content material.