The most consequential AI announcement of June 2026 might be the most boring-looking one. No new frontier model, no benchmark, no demo that makes the rounds on social media. Just a specification — a plain-Markdown convention for how AI agents read your company's curated knowledge. Google Cloud calls it the Open Knowledge Format, or OKF, and the temptation is to file it under “developer plumbing” and move on.
Don't. The single most expensive mistake mid-market companies are making with AI right now is letting their hard-won institutional knowledge get trapped inside one vendor's proprietary format — so that the day they want to switch models or platforms, switching means rebuilding everything from scratch. OKF matters far less as a Google product than as a signal that something important is becoming a buying criterion: portable, model-agnostic context that you own. That's the argument I want to make here.
Key Takeaways
- OKF is a vendor-neutral, plain-Markdown spec for giving AI agents curated business context — a directory of markdown “concept” files with YAML frontmatter, no proprietary SDK or runtime required.
- The real story is the direction: portable, model-agnostic context is becoming a procurement requirement, not a nice-to-have.
- Vendor lock-in is a concrete, expensive risk — surveys put a large majority of organizations on record as worried about it, and migrations routinely run into six and seven figures.
- The fix is owning your knowledge in an open, human-readable, portable format that works against today's model and whatever wins tomorrow.
- This is distinct from “how to structure a knowledge base” — it's about who owns the format, and why neutrality is leverage you can keep.
What Is the Open Knowledge Format, Exactly?
OKF v0.1 is an open specification that standardizes how AI agents access an organization's curated knowledge. As MarkTechPost reported, it represents that knowledge as a directory of Markdown files with YAML frontmatter — no proprietary SDK, runtime, or registry required to read it.
The structure is deliberately minimal:
- Each file is a “concept.” A concept can be a table, a dataset, a metric, a runbook, or an API. One concept per file, and the file path serves as its identity.
- Frontmatter carries the metadata. Reserved YAML fields include
type,title,description,resource,tags, andtimestamp. The only strictly required field istype. - Markdown links form a knowledge graph. Cross-references between concept files create a navigable web of related knowledge.
- Optional
index.mdandlog.mdprovide a directory of contents and a change history.
The design principles are the whole point. OKF is, in its own framing, “a format, not a platform” — minimally opinionated, vendor-neutral, and built so that producer and consumer are independent. A bundle can be hand-written by a person or generated by a pipeline, and either way it should work with any tool that understands the format. Google Cloud shipped some reference tooling around it — a BigQuery enrichment agent, a static HTML visualizer, sample bundles — but none of that is required to use the spec.
It also draws an explicit line against the default approach. Where retrieval-augmented generation (RAG) derives answers from raw document chunks at query time, OKF stores curated, version-controlled, directly-editable concepts that an agent reads — and can update. That's an architectural choice with real consequences, and it's not a new instinct.
Where This Idea Came From: Karpathy's LLM Wiki
OKF didn't appear from nowhere. MarkTechPost notes it was inspired by Andrej Karpathy's “LLM Wiki” concept from April 2026. In his original write-up, Karpathy described a pattern where, instead of pointing a model at raw documents and retrieving chunks, you let an agent incrementally build and maintain a persistent wiki of Markdown files — immutable raw sources in one layer, an LLM-generated wiki of summaries and concept pages in another, and a schema file telling the agent how to ingest, query, and maintain it.
The striking part was scale: Karpathy reported his own wiki grew to roughly 100 articles and 400,000 words, and at that size the model could still navigate it efficiently using the index and summaries — faster and more accurate than a RAG pipeline for his use case. The labor split is what makes it work: the human curates sources and asks good questions; the model does the persistent bookkeeping of summarizing, cross-referencing, and filing.
We've written before about the architectural side of this — the shift from chunk-and-retrieve toward a curated, compounding knowledge layer — in our piece on knowledge base architecture beyond RAG. That article is about how to structure the knowledge. This one is about something the structure alone doesn't settle: who owns the format it lives in. OKF's contribution is making “your knowledge is portable Markdown you own” into a named, shared standard rather than a private convention.
Why Is Vendor Lock-In the Real Story?
Here's the uncomfortable arithmetic. The context that makes an AI Employee genuinely useful — your SOPs, product data, pricing logic, policies, the accumulated “how we actually do things here” — is the most valuable and least portable asset in your AI stack. And the more of it you pour into one vendor's proprietary system, the harder and more expensive it becomes to ever leave. That's not a hypothetical; it's the dynamic security and infrastructure analysts now call data gravity.
The concern is widespread and on the record. The Register reported in April 2026 that AI vendor lock-in is already biting enterprise budgets. A 2026 enterprise guide from Swfte compiles the survey picture: 67% of organizations aim to avoid high dependency on a single AI provider, 45% say lock-in has already hindered their ability to adopt better tools, and 84% now factor digital sovereignty into their AI strategy. That same guide reports a migration typically costs about twice the initial investment, and that 57% of IT leaders spent more than $1 million on platform migrations in a single recent year.
The cost isn't abstract. StackAI's breakdown of hidden lock-in costs lays out where the money goes when you try to leave: data egress fees, engineering rewrites for proprietary SDK dependencies, dual-run infrastructure during the cutover, and revalidation overhead. Its blunt summary is that lock-in is “often a six-figure cost event even for a single workload, and a seven-figure event for shared platforms.”
And the pace of model change makes this worse, not better. The Swfte data shows just how fast the leaderboard reshuffles — in its figures, one provider's share moved from 12% to 40% across roughly two years while another fell from 50% to 27%. If the best model for your use case might be a different vendor's in eighteen months, then anything that makes switching painful is a tax on your ability to use the strongest AI available. This is the same flexibility-versus-capture tension that Kai Waehner's 2026 enterprise agentic AI landscape frames as one of the defining strategic questions of the year.
What Portable Context Buys You
So what's the strategic move? Treat your knowledge format the way you already treat your data backups and your domain name: as something you own outright, in a form no single vendor controls.
Concretely, a portable, open-format knowledge base gives you four kinds of leverage:
| What you get | Why it matters |
|---|---|
| Model optionality | The same Markdown context works against Claude today and whatever model wins tomorrow — you switch the engine, not the knowledge. |
| Negotiating leverage | A vendor that knows you can leave prices and behaves differently than one that knows you can't. |
| Auditability | Human-readable Markdown is something your team can actually read, review, and correct — not a black-box embedding store. |
| Resilience | If a vendor sunsets a product or changes terms, your institutional knowledge survives the transition intact. |
This is why I keep arguing that neutrality is leverage. We made the broader version of this case in our look at building a neutral evaluation layer for multi-model AI, and in our analysis of model sovereignty and procurement leverage. A standard like OKF is a concrete, adoptable expression of that principle: it turns “we should avoid lock-in someday” into “our knowledge already lives in a format anyone can read.”
It also reinforces something we've argued about the difference between generic AI tools and custom AI Employees: the thing that makes an AI Employee actually good at your business is your curated context. If that context is the source of the value, it should be the thing you own most firmly — not the thing most tightly bound to a vendor you might outgrow.
How to Adopt This Without Betting on One Spec
To be clear and honest about the trade-offs: OKF is at v0.1, it's young, and ecosystem support is early. I am not telling you to rip out what you have and “go OKF” this quarter. Standards at this stage can evolve, merge, or get superseded. Betting your operations on a v0.1 spec would be its own kind of risk.
The durable principle survives whichever specific standard wins:
- Author your knowledge in open, human-readable formats. Markdown with structured frontmatter is readable by people, diff-able in version control, and convertible to almost anything. That's true whether or not you formally adopt OKF.
- Keep raw sources separate from generated summaries. The Karpathy-style split — immutable sources in one layer, an AI-maintained wiki on top — means you can always regenerate the derived layer for a new model.
- Version-control the whole thing. Your knowledge base belongs in a repository you control, with history, not in a vendor's opaque store.
- Put it behind your own gateway. Portable context is only an asset if it's also governed. We covered the architecture for that in our piece on running context behind a secure AI gateway — ownership and security are two sides of the same decision.
Adopt the posture now — portable, owned, auditable, governed — and you can adopt the specific format when the standard matures, with little wasted work either way.
A Note for Northeast Indiana Operators
For the Fort Wayne and Northeast Indiana businesses we work with, the lesson doesn't require a standards committee to appreciate. If you're a regional manufacturer, accounting firm, or home-services company investing real effort into making an AI tool understand your business, ask one question before you go deeper: if this vendor disappeared tomorrow, what would I keep? If the answer is “a pile of context locked in their system,” you've built your most valuable AI asset on rented land. Keeping your knowledge in a format you own is the unglamorous insurance policy that keeps your options open as the AI market keeps reshuffling.
Make it concrete. Say a DeKalb County manufacturer spends six months teaching a chatbot its part numbers, quoting rules, and the tribal knowledge of which jobs always run late. If all of that lives only inside one vendor's account, the company hasn't built an asset — it's renting one, and the rent can go up. Authored as plain Markdown the company keeps in its own repository, that same body of knowledge is portable: it moves to a new model, survives a price hike, and can be read and corrected by the office manager who actually knows the business. For a mid-market team without a platform-engineering bench, that ownership is the difference between switching vendors in an afternoon and being quietly held hostage by the one you started with.
The Bottom Line
OKF probably won't trend. But it names something that should be a non-negotiable in every mid-market AI decision from here on: your institutional knowledge should live in an open, portable, model-agnostic format that you own and can read. The models will keep changing. The vendor leaderboard will keep reshuffling. The one thing that should stay yours is the curated context that makes any of them useful for your business.
At Cloud Radix, we build AI Employees on exactly that principle — your knowledge in a portable, owned format, behind our Secure AI Gateway, working against the best model for the job rather than the one you happen to be locked into. If you want to deploy AI without signing away your most valuable asset, let's talk about a vendor-neutral approach.
Frequently Asked Questions
Q1.What is the Open Knowledge Format (OKF)?
OKF is a vendor-neutral, open specification from Google Cloud for giving AI agents curated organizational context. It represents knowledge as a directory of Markdown files with YAML frontmatter — each file a "concept" like a metric, runbook, or dataset — with no proprietary SDK or runtime required to read it. The only strictly required frontmatter field is type.
Q2.How is OKF different from RAG?
Retrieval-augmented generation derives answers from raw document chunks at query time, with no curated, stable representation of your knowledge. OKF instead stores curated, version-controlled, directly-editable concept files that an agent reads and can update. It's the difference between re-deriving knowledge on every query and maintaining a deliberate, owned knowledge layer.
Q3.Why does AI vendor lock-in matter for mid-market companies?
Because the context that makes an AI useful — your SOPs, data, and policies — is your least portable asset, and switching costs are real. Industry guides report that the majority of organizations are actively trying to avoid single-vendor dependency, that lock-in has already blocked many from adopting better tools, and that migrations frequently run into six and seven figures. As models keep changing, anything that makes switching painful taxes your ability to use the strongest AI available.
Q4.Should I adopt OKF right now?
Adopt the principle now; adopt the specific spec when it matures. OKF is at v0.1 and the ecosystem is young, so betting operations on it today carries its own risk. The safe move is to author your knowledge in open, human-readable Markdown with structured frontmatter, keep raw sources separate from AI-generated summaries, version-control it, and put it behind your own gateway. Those choices pay off regardless of which standard ultimately wins.
Q5.Is a portable knowledge base less secure than a vendor's system?
Not inherently — ownership and security are separate decisions that should be made together. A portable, human-readable knowledge base is actually easier to audit because your team can read and review it. The key is to govern access through your own secure AI gateway rather than scattering the data, so that portability and control reinforce each other instead of trading off.
Q6.What does "neutrality is leverage" mean in practice?
It means a vendor that knows you can leave treats you differently than one that knows you can't. If your knowledge lives in an open format that works across models, you keep model optionality, negotiating power, and resilience against a vendor changing terms or sunsetting a product. Lock-in quietly transfers that leverage to the vendor; portability keeps it with you.
Q7.How should a Northeast Indiana business start without overcommitting to OKF?
Begin with the posture, not the spec. Pick one well-understood process — your quoting rules, your onboarding checklist, your most-asked customer questions — and write that knowledge as plain Markdown files with simple structured headers, kept in a version-controlled repository your team controls. That gives you something portable and auditable today, regardless of which AI tool reads it. For Fort Wayne and DeKalb County operators, this is also a low-risk first project: small enough to finish in a week, valuable enough that the AI Employee built on it pays for itself, and owned outright so nothing is wasted if you change vendors or the standard evolves.
Sources & Further Reading
- MarkTechPost: marktechpost.com/2026/06/16/google-cloud-introduces-open-knowledge-format-okf — Google Cloud introduces Open Knowledge Format (OKF), a vendor-neutral Markdown spec for giving AI agents curated context.
- Andrej Karpathy (GitHub Gist): gist.github.com/karpathy/442a6bf555914893e9891c11519de94f — llm-wiki, the original LLM Wiki concept that inspired OKF.
- Swfte AI: swfte.com/blog/avoid-ai-vendor-lock-in-enterprise-guide — AI Vendor Lock-in: How Enterprises Are Breaking Free in 2026.
- StackAI: stackai.com/insights/the-hidden-costs-of-vendor-lock-in-for-ai-infrastructure — The Hidden Costs of Vendor Lock-In for AI Infrastructure.
- The Register: theregister.com/software/2026/04/28/locked-stocked-and-losing-budget-ai-vendor-lock-in-bites — Locked, stocked, and losing budget: AI vendor lock-in bites.
- Kai Waehner: kai-waehner.de/blog/2026/04/06/enterprise-agentic-ai-landscape-2026 — Enterprise Agentic AI Landscape 2026: Trust, Flexibility, and Vendor Lock-in.
Deploy AI Without Signing Away Your Most Valuable Asset
We build AI Employees on a portable, owned knowledge base behind our Secure AI Gateway — so you keep model optionality, negotiating leverage, and your institutional knowledge no matter how the market reshuffles.
Talk About a Vendor-Neutral ApproachServing Fort Wayne, Auburn, and Northeast Indiana. No lock-in — just an honest conversation.
