# moca-reader-langchain

> moca-reader-langchain: a LangChain retriever over MOCA packages, returning citation records as Documents, with trust rules applied.

Source: https://openmoca.org/readers/python/langchain/

A [LangChain](https://python.langchain.com) retriever over MOCA packages,
built on the [MOCA Reader for Python](/readers/python/).

```sh
pip install moca-reader-langchain
```

```python
from moca_reader import LexicalBackend, Library, Search, read_package
from moca_reader_langchain import MocaRetriever

pkg = read_package("support-kb.moca", trust_root="trust-root.json")
library = Library().add(pkg)
search = Search(library, LexicalBackend(library), audiences=["public"])

retriever = MocaRetriever(search=search, limit=3)
for doc in retriever.invoke("How long is the refund window?"):
    print(doc.metadata["moca_node_ref"], doc.metadata["moca_score"])
    print(doc.page_content[:80])
```

`MocaRetriever` wraps the `Search` you build: its backend (lexical, dense,
hybrid or your own store), its audiences and its hooks. Every call goes through
the Reader's gate, so content that is out of force, superseded, deprecated or
outside your audiences never comes back, and nothing a caller passes can widen
that. `limit`, `locale`, `concepts`, `scope` and `include_all` are set when you
build the retriever.

Each result is a LangChain `Document`:

| Field | Holds |
| --- | --- |
| `page_content` | The text of the node, or of a sidecar chunk |
| `id` | The node's versioned reference and byte span |
| `metadata["moca_citation"]` | The whole citation record: package, version and digest, trust signals, evidence |
| `metadata["moca_*"]` | The flat fields of the consuming guide (`moca_node_ref`, `moca_package_digest`, `moca_status`, `moca_signed`, ...), for vector stores that need flat metadata |
| `metadata["moca_score"]` | The search score |

**Package text is untrusted input.** Put `page_content` in a prompt as quoted
reference material to cite, never as instructions
([Reader contract §12](/spec/reader-contract/#12-handing-content-to-a-model)).
Show users the source (`moca_node_ref`, the package version) with each answer.

The async path (`ainvoke`) runs the same search.

## Licence

Apache-2.0.