MOCA and RAG

Retrieval-augmented generation (RAG) grounds a model’s answer in content retrieved for the question. Most RAG pipelines start by flattening knowledge into chunks in a vector store. On the way, the knowledge loses what made it trustworthy: which version it was, who published it, whether it is still in force, who it was for, how it was organised, and the source each statement came from.

MOCA keeps those things with the knowledge, so RAG can use them. A MOCA package is the knowledge a RAG pipeline retrieves from, with its version, its structure and its original sources, sealed by one digest. A MOCA Reader does the retrieval, or sits in front of your own vector store, and every result comes back as a citation that can be checked.

What changes in a RAG pipeline

RAG problem	Without MOCA	With MOCA
Citations nobody can check	The answer names a document, if anything	Every result is a citation record: the passage, a versioned reference to the page, and the quote it relies on, marked matched when found in the original inside the package
Stale or superseded content	Old chunks stay in the index until someone rebuilds it	Pages and packages carry validity dates, and a newer package can supersede an older one; the Reader leaves out content that is not in force or has been replaced, at search time
Content for the wrong audience	Internal pages leak into public answers unless every pipeline filters them	Each page has an audience; the host’s audience filter runs on every search, and a caller cannot widen it
Lost structure	Steps out of order, prerequisites missing	An optional ontology keeps parts, order and prerequisites, and search can add what a result requires (ontology)
Unknown provenance	Nobody can say who published or checked the content	Packages are signed by their publisher, pages by their reviewers, verified against your trust root
Locked to one stack	Re-chunk and re-embed for every model, vendor and framework	One package, read the same way by the TypeScript, .NET and Python Readers, an MCP server, or your own vector store
Proving what was read	Logs of chunk ids from an index that has since been rebuilt	Log the package digest and the versioned page reference with each answer: they name the exact text, for good

Where MOCA sits

flowchart LR
  pkg["MOCA package<br/>pages · structure · sources"] -- "read and verify" --> reader["MOCA Reader<br/>search · trust rules"]
  pkg -. "moca-index" .-> idx["Sidecar index<br/>optional<br/>lexical or semantic"]
  idx -- "bound to the digest" --> reader
  reader <-->|"ingest · search"| store[("Your vector store<br/>optional")]
  reader --> cite["Citation records"] --> llm["Your model<br/>and application"]

1. The Reader reads and verifies the package: its digest, signatures and evidence.
2. It searches in one of three places, through the backend your application chooses: the package’s own text (lexical), a sidecar index built from the package with moca-index (lexical, semantic or hybrid), or your own vector store, which you fill at ingest and the Reader searches at query time.
3. The same trust rules run on every result, wherever it was found, and each comes back as a citation record for your model and application.

You keep your model, your framework and, if you have one, your vector store. MOCA adds the knowledge layer underneath and the trust rules in the middle:

• Search over the package itself, with lexical, semantic or hybrid search (search and sidecar indexes). Semantic search uses a sidecar index of vectors, built once per package version.
• Or search your own vector store through a store backend: verify each package at ingest, write each chunk with its citation record, and the Reader applies the same trust rules to every hit (consuming packages).
• Or serve it over MCP, so any MCP client can search and cite it.

RAG frameworks

• Python: moca-reader-langchain and moca-reader-llamaindex are retrievers that return citation records as LangChain Documents and LlamaIndex nodes.
• .NET: OpenMoca.Reader.AgentFramework grounds a Microsoft Agent Framework agent in packages, and OpenMoca.Reader.VectorData searches any Microsoft.Extensions.VectorData store.
• Any language: the MOCA MCP server gives any MCP client search, structure and citations.

What MOCA does not do

It does not choose your model, embed for you, or control what a model says. It makes sure the model is given the right content, and that every passage can be traced to its source. Treat package text as quoted reference material, never as instructions (Reader contract §12).

Try it

Try OpenMOCA runs lexical, semantic and hybrid search over eight real packages, and shows where each result ranked in each list. The walkthrough builds and serves a package to an AI agent in fifteen minutes.