Use cases

MOCA is not another model, vector store or agent framework. It is the form your knowledge travels in, so that whatever you build (a RAG pipeline, an AI agent, a search page) answers from the right version, cites the original source, and can prove both.

It pays off when knowledge has to keep its structure and its sources wherever it is read, or when it crosses a boundary: between organisations, between teams, or between the people who own it and the systems that serve it.

Every use case below gets the same foundations:

• One digest seals the package, and a signature shows who published it.
• Every answer is a citation to a versioned node, with a quote checked against the original source.
• Trust rules apply at search time. Content that is out of force, superseded or for another audience never comes back, whatever search backend you use.

At a glance

Use case	What MOCA adds	See it
Sharing knowledge for grounding	Package a body of knowledge once; any assistant, model or framework grounds on it	Try it
Manuals and procedures	Steps in order, the checks that come first, cited to the PDF page, offline	Try it
Law and regulation	Only the version in force, with the exact clause cited	Try it
Courses and training	Modules, lessons in order and prerequisites, cited to the moment in a lecture	Try it
Recorded talks and video series	Searchable speech, every answer opening the video at the moment	Try it
Onboarding	Foundations before advanced material, with named owners’ reviews	Try it
Documentation shipped to customers	Customers verify who published it, and stop quoting last year’s release	Try it
Customer support knowledge	AI drafts kept apart from what a person checked, internal answers kept internal	Try it
Audit and regulated AI	A record of exactly which content each answer used	Try it
Clinical guidelines	The pathway’s order kept, every citation checked against the guideline

These are examples, not a complete list: MOCA fits wherever the signs above hold. They are described at the same depth on purpose. MOCA is neutral between domains, and none of them is special to the standard. The “Try it” links open a package in Try OpenMOCA, an example site you can search without an account.

Sharing knowledge for grounding

The problem. A body of knowledge, such as a team’s documentation, a project’s docs or a published guide, has to be loaded again into every assistant, model and pipeline that should answer from it. Each copy is chunked and indexed its own way, loses its structure and its sources, and drifts from the original. Nobody can say which version an answer came from.

What MOCA adds.

• Package once, ground anywhere. The same .moca file is read in-process by the TypeScript, .NET or Python Reader, served to any MCP client, or loaded into your own vector store, with the same trust rules and citation records each time. Changing model or framework does not mean rebuilding the knowledge.
• Start small. A package can be just a manifest and a folder of Markdown, with no structure and no signatures. That is enough to move a grounded project from one assistant to another.
• Grow when you need to. Add structure, evidence, a sidecar search index or a signature later, without changing how anyone reads the package.
• Every answer is traceable. A citation names the node, its version and the package digest, whoever answered.

See it. examples/minimal and the quickstart: a valid package in five minutes. Try the Kubernetes documentation: all 1,653 pages in one package.

Manuals and procedures

The problem. A technician needs the current procedure, in the right order, often with no connection. A chatbot over a folder of PDFs returns a step out of order, or one from a withdrawn revision, and nobody can tell which page it came from.

What MOCA adds.

• The order and the checks are kept. Structure holds the assembly tree (dcterms:hasPart), the steps in order (skos:OrderedCollection) and the checks that come first (dcterms:requires). Every Reader answers “what must happen before step 5?” the same way.
• Answers stay on the job. Search can be limited to the system being repaired and everything under it.
• Withdrawn procedures stop being used. A new revision supersedes the old one, and the Reader leaves the old one out.
• It works offline. A .moca file with a sidecar index runs on a laptop or tablet. Lexical search needs nothing more; semantic search needs only a small local embedding model.
• Your PDFs are the starting point. moca-convert splits each PDF into a node per section, and evidence cites the pages each node came from (page=12, with the printed page number).
• Large manuals are composed. A handbook can be built from chapter packages, each pinned by its digest, so the handbook fixes exactly which chapter versions it contains.

See it. examples/pdf, the FAA’s 19-PDF flying handbook, and examples/handbook, a handbook composed of chapters. Try the Airplane Flying Handbook, where each topic shows what it builds on, and the HSE guidance, cited to the PDF page.

Law and regulation

The problem. “Which clause applies, and what does it depend on?” has one right answer. A system that mixes last year’s text with this year’s, or paraphrases without a citation, gives a confident wrong one.

What MOCA adds.

• Only what is in force. Each version is a package with a validity window. The default retrieval policy leaves out versions that are not in force, and supersedes retires them without deleting them, so “what did it say in 2025?” still has an answer.
• The shape of the law is kept. Structure holds the part, section and clause tree, cross-references (dcterms:requires) and amendments (dcterms:replaces).
• The exact clause is cited. Evidence points at the clause in the original text, and the Reader marks it matched when the quote is found there.

See it. examples/policy-corpus, two versions of one policy. Try the Data Protection Act 2018, section by section.

Courses and training

The problem. A course has modules, a sequence and prerequisites. An assistant that ignores them answers with material the learner has not reached, or cites “the lecture” with no way to find the moment.

What MOCA adds.

• The course’s shape is kept. Structure holds modules as parts, lessons in order, and prerequisites as requires. An application can use them to keep a learner on their current module.
• Citations to the moment. Evidence can point at a time range in a lecture (Media Fragments). The WebVTT captions travel in the package, so the citation is checked against what was said.
• Your teaching vocabulary stays yours. Learning objectives and similar terms travel as an overlay or extra vocabulary, never as part of the standard.

See it. Try Learn Copilot Studio 2026, a 17-video series in order.

Recorded talks and video series

The problem. A talk or a video series is knowledge locked in speech. You cannot search it, and an answer about it cannot point to where it was said.

What MOCA adds.

• Speech becomes searchable text. Each video is a node holding its description and a transcript from its captions. The video stays where it is published, as the node’s resource.
• Every answer opens at the moment. Evidence cites chapters as time ranges, and the captions travel under media/, so every range is checked without a network.
• A series stays in order (skos:OrderedCollection).
• Nothing is written by hand. moca-youtube builds the package from a video or a playlist.

See it. examples/youtube. Try First Steps to Using Claude AI, a recorded talk cited to its chapters.

Onboarding

The problem. A new starter should meet the foundations before the advanced material, and the handbook they learn from should be one someone actually checked.

What MOCA adds.

• Foundations first. Structure holds skill areas, the learning path, and foundations as requires.
• Your own rules, without editing the package. An overlay adds the organisation’s rules, for example a compliance module required before system access.
• Named owners. Review attestations show which pages a named owner has checked, and when.

See it. Try the GitLab handbook, a company’s policies and people practices.

Documentation shipped to customers

The problem. A software vendor’s customers load its documentation into their own assistants. Copies drift, last year’s behaviour keeps being quoted, and nobody can tell an official copy from a modified one.

What MOCA adds.

• Customers verify the publisher. They check the publisher attestation before loading the package.
• Old releases stop being quoted. Each release is a new version that supersedes the last.
• It fits the customer’s stack. Whatever Reader, MCP client or vector store the customer uses, the same trust rules apply.
• The exact paragraph is cited, in the documentation or the release notes.

See it. Try the Kubernetes documentation, vendor documentation at scale.

Customer support knowledge

The problem. Support answers drift, and AI drafting tools rewrite them. Soon nobody knows which answers a person has checked, and internal procedures leak into the public assistant.

What MOCA adds.

• Drafts are marked. OKF generated marks drafted content, and a support lead’s review attestation marks what a person actually checked.
• Answers are re-checked. stale_after sets when a node is due for review again.
• Internal stays internal. audience: internal keeps agent-only procedures out of the public assistant, before retrieval.
• One answer, several languages. A node can carry translations, and the Reader picks the one for the caller’s language.

See it. examples/support-kb. Try the Acme Support Knowledge Base: signed, reviewed, in two languages, with superseded content flagged.

Audit and regulated AI

The problem. An organisation must show which knowledge its assistant used for an answer, and where that knowledge came from. Logs of “retrieved chunk 4711” from a vector store that has since been rebuilt prove nothing.

What MOCA adds.

• A record that names the exact content. Store the versioned node reference and the package digest with every answer. Together they name the exact bytes the answer came from, long after the index is rebuilt.
• The sources are kept. The originals travel under sources/, so each citation can be matched against its source at audit time.
• A clear scope. Dataset governance records, such as an SPDX 3.0 Dataset or Croissant document, sit beside the package, and risk classification stays with the AI system, not the package.

See it. Try inspecting a package: its digest, its signer, the Reader’s findings and its search indexes.

Clinical guidelines

The problem. A guideline is a pathway: tests before treatments, conditions and their subtypes. Retrieval that ignores the pathway can surface a treatment before the test it depends on.

What MOCA adds.

• The pathway is kept. Structure holds the conditions (skos:broader), the pathway stages in order, and what each stage requires.
• Every citation can be checked. The original guideline travels inside the package, so a citation is matched against it with no network.
• A clear limit. MOCA grounds an assistant over the guideline. It does not replace clinical decision-support standards, and structure narrows what is retrieved, not what a model says.

Where MOCA is a poor fit

• Live operational data, such as tickets, stock levels or metrics. Query the system of record.
• One team, one pipeline, and no one asking where an answer came from. Use OKF on its own.
• Anything that needs configuration to travel with content. Packages never configure their Readers.