RAG
===

Retrieval-augmented knowledge lives in Django app ``agency.corpus``.

Shape
-----

::

   Tenant → Workspace → Document → (extraction) → Chunk → Chroma collection

Workspaces carry embedding / chunk defaults and deterministic collection names
(per tenant + workspace). Agents can query private knowledge without mixing that
infra into ThreadRunner itself.

In this **reference** stack, vectors are stored in **ChromaDB**. That is a host
deployment choice; the package talks to the configured store.

API surface (today)
-------------------

REST viewsets and actions under ``/api/`` include workspace and document
resources (search / ask / reindex-style actions as implemented on the viewsets).
See :doc:`framework/agency/api` and OpenAPI at ``/api/``.

Related: **miner** (``Scrap`` / ``Block``) is scrape ingest, not the RAG index —
:doc:`framework/agency/miner`.

Module reference
----------------

.. automodule:: agency.corpus
   :members:
   :undoc-members:

.. automodule:: agency.corpus.models
   :members:
   :undoc-members:
   :show-inheritance: