01 — The problem
What has to stop going wrong.
You want an assistant over your own data, and you cannot have it inventing figures or reading records the user is not allowed to see.
02 — What we build
The system behind the surface.
Retrieval scoped to the caller's permissions, a whitelist of tools the model selects from rather than composes, and a grounding check that every number it states appeared in a real result. The model never writes a query.
01
Bound it
Define the failure modes, permissions and consistency rules before the interface hides them.
02
Build it
Ship production code in your repository, in phases with a written exit state.
03
Try to break it
Test the paths that carry money, identity, isolation and irreversible decisions.
03 — Decision material
Know whether it fits before a call.
A strong fit
- Teams searching internal documents or policies
- Products that need answers from live operational records
- Workflows where permissions must survive retrieval
What leaves with you
- Corpus, permissions and failure-mode map
- Ingestion, retrieval and citation pipeline
- Held-out evaluation set and refusal tests
- Monitored deployment and handover
04 — Evidence
A claim should lead somewhere.
An assistant that answers from a business's own records and cannot state a figure that is not in them.
Shown in Nexora AI — Retail and trading