Can you check where your AI’s answer came from?
- Which version said it
- Who published it
- Whether anyone checked it
- Whether it still applies
Knowledge for AI is often locked in, and hard to check
Locked in one pipeline
Chunked, embedded and stored for one product.
Often rebuilt for each model, vendor and pipeline.
Hard to check
Structure and sources are often lost on the way in.
Few answers say which version of which page they rest on.
Hard to prove
Changed since it was approved?
Still in force? Superseded?
Most pipelines can’t tell you.
So how do we resolve this?
Package the knowledge,
not the pipeline.
Not another model, vector store or agent framework.
An open standard for portable grounding for AI.
Modular
the package
Build it once, use it anywhere.
Not locked into one pipeline.
Combine packages, pinned by digest.
Swap search indexes per model.
Ontology
the structure layer
Concepts, hierarchy, order and what requires what.
W3C SKOS at its core.
Search stays in scope.
Richer models, like OWL, travel with it.
Optional
Content
the content and evidence
One Markdown node per topic.
Validity windows, audiences and translations.
Every answer points back to its source.
Open Knowledge Format
Attestation
the trust through all three
One digest covers every file, so any change shows.
Signatures prove who published it, and who checked it.
Signing optional
Where it helps
Sharing knowledge for grounding
Package a body of knowledge once; any assistant, model or framework grounds on it.
In the demo: Kubernetes, 1,653 pages
Manuals and procedures
Steps in order, the checks that come first, cited to the PDF page, even offline.
In the demo: FAA handbook, HSE guidance
Law and regulation
Only the version in force, with the exact clause cited.
In the demo: Data Protection Act 2018
Courses, talks and video
Lessons in order and prerequisites kept; every answer opens the video at the moment.
In the demo: a course and a talk
Customer support knowledge
AI drafts kept apart from what a person checked; internal answers stay internal.
In the demo: Acme Support
Audit and regulated AI
A record of exactly which content each answer used, long after the index is rebuilt.
In the demo: every citation record
DEMO 1
Try OpenMOCA
try.openmoca.org: real packages, in the browser, no account
What you just saw
- Search, with no accountLexical, semantic or hybrid, side by side, with no AI model.
- Open every result at its sourceThe passage marked in place; a video opens at the moment.
- Use the structureSearch within a topic, and see what it builds on.
- Check, don’t trustThe digest, the signer, the findings, the indexes.
- Ask it questionsEvery answer cites the package; a citation opens the source.
A package, in three layers
.mocaA set of files with moca.json at its root: a folder, a .moca archive (a Zip), or records a host supplies. All give the same digest.acme-support/ moca.json id, version, licence content/ the pages sources/ the originals structure.ttl the ontology attestations/ signatures, reviews
.moca.idxA folder or archive (a Zip), kept outside the package it indexes. Optional; built for one package version and digest.What a Reader gives you
1. Verify
The digest, the signatures against your trust root, and every finding.
2. Search
Lexical, semantic or hybrid, in one call, with the trust rules applied.
3. Cite
Every result is a citation record: the version, who signed and reviewed it, and its evidence.
Three open-source Readers, one conformance corpus:
a package reads the same way everywhere.
Search where your data already is
A local index
A sidecar beside the package: words, plus vectors from four local models.
Built for one package version and digest. Rebuild it any time.
Offline: no AI provider sees a thing.
Your vector store
Verify each package as you load it in. The Reader searches the store.
The trust rules apply to every result.
VectorData (.NET), LangChain, LlamaIndex
Hybrid
Keyword and meaning together, merged by rank.
Reciprocal rank fusion you can tune: k, and a weight for each side.
One search call, whichever backend.
Defaults: k 60, equal weights
Every answer comes with a citation record
{
"package": {
"version": "1.0.0",
"digest": "sha256:3fb4f2bb…",
"signed": true
},
"node": {
"title": "Refund window",
"ref": "…/acme-support@1.0.0#returns/refund-window.md"
},
"trust": {
"inForce": true,
"superseded": false,
"attestedReviews": [
{ "signer": "Sam Ortiz", "outcome": "accurate" }
]
},
"evidence": [{
"source": "Acme Customer Terms, section 7",
"exact": "…a full refund within 30 days of delivery.",
"matched": true
}]
}
Shortened. Each search result from a Reader is a citation record: openmoca.org/spec/reader-interface
- Which versionThe package version and digest, and the page.
- Who stands behind itSigned by Acme, reviewed by Sam.
- Is it still trueIn force, and not superseded.
- The sentence it rests onQuoted from the source, and checked.
Open tools, from a folder to a signed package
Convert
Markdown, Obsidian, PDF, OpenAPI, YouTube
moca-convertCheck
As a Reader would.
Pack to .moca.
moca-lintSign
The publisher, and reviews of each page.
moca-signIndex
Lexical and semantic, with local models.
moca-indexServe
To any MCP client, on your computer.
moca-mcpThe tools run with npx, nothing to install. The MCP server is one download.
Or let your AI assistant run them: the openmoca agent skill teaches it every tool.
Your own assistant, over MCP
- Runs on your computerNo network calls.
- Checked firstThe digest and signatures, every load.
- Only your audienceInternal and expired pages stay out.
Claude Desktop: Settings → Extensions → Edit Config
{
"mcpServers": {
"moca": {
"command": "/Users/you/moca/moca-mcp",
"args": [
"--library", "/Users/you/moca",
"--trust-root",
"/Users/you/moca/example-trust-root.json",
"--audience", "public"]
}
}
}
Claude Code: one line
claude mcp add moca -- ~/moca/moca-mcp \ --library ~/moca \ --trust-root ~/moca/example-trust-root.json \ --audience public
VS Code: .vscode/mcp.json, or MCP: Add Server
{
"servers": {
"moca": {
"type": "stdio",
"command": "/Users/you/moca/moca-mcp",
"args": [
"--library", "/Users/you/moca",
"--trust-root",
"/Users/you/moca/example-trust-root.json",
"--audience", "public"]
}
}
}
Then ask in Copilot Chat, in agent mode.
DEMO 2
Your own assistant
Claude Desktop, with the MOCA MCP server
Works with what you already use
Make it from
- A folder of Markdown
- An Obsidian vault
- PDFs
- OpenAPI documents
- YouTube videos
Use it in
- Claude Desktop and Claude Code
- VS Code, with GitHub Copilot
- Cursor
- JetBrains IDEs, Cline
- Any local MCP client
Build with
- TypeScript, .NET, Python
- Microsoft Agent Framework
- Microsoft.Extensions.AI
- LangChain
- LlamaIndex
The agent skill: ask Claude, GitHub Copilot or Codex to make, sign, search and cite packages for you.
Make your own package: four commands
npx @openmoca/moca-convert ./docs -o my-docs \ --id https://example.com/moca/my-docs --title "My docs"
npx @openmoca/moca-lint lint my-docs
npx @openmoca/moca-sign sign my-docs --key publisher.pem --keyid publisher
npx @openmoca/moca-index build my-docs -o my-docs.moca.idx \ --model bge-small-en-v1.5
Then use it: in your app, or in your assistant over MCP
npm install @openmoca/moca-core
dotnet add package OpenMoca.Reader
pip install moca-reader
moca-mcp: one download
THANK YOU
Questions?
- openmoca.orgThe spec, guides for each Reader, the tools and the walkthrough.
- try.openmoca.orgSearch real packages, and open every result at its source.
- github.com/openmocaThe spec, the Readers, the tools and the MCP server. Apache 2.0.
Slides: openmoca.org/talks
