Skip to content
AMPHOZE
DemonstrationRegulated services

Enterprise Knowledge Intelligence

A retrieval and agent architecture for internal knowledge where permissions, provenance and refusal behaviour are requirements rather than refinements.

Industry
Regulated services
Scenario
Illustrative — regulated services organisation
Status
In-house demonstration

This is an illustrative example of the kind of work we do, not a client engagement. The scenario is written to make the engineering decisions legible; the outcomes below describe what the system was designed to do, not verified client metrics.

Challenge

The constraint we designed against.

Internal knowledge is scattered across a wiki, a document store, ticket history and mail, with permissions that differ per source and matter legally. A naive index over all of it is a data-leak mechanism with a chat interface. The harder problem is not retrieval quality but knowing when the system should decline to answer, and proving afterwards what it read.

Solution

What we built, and why it is shaped this way.

Documents are chunked with structure preserved and indexed with their source access-control list attached, so retrieval filters by the caller identity before ranking rather than after. Hybrid lexical and vector search feeds a reranking pass, and answers are assembled only from retrieved spans, each citing its source. The system is built to refuse: below a support threshold it says it does not know, which is the behaviour a regulated environment actually needs. A fixed evaluation set — answerable, unanswerable and permission-trap questions — is graded on every index, prompt or model change, and every query logs the identity, the retrieved span identifiers and the returned citations.

Technology

The stack behind it.

  • Python
  • TypeScript
  • Postgres
  • pgvector
  • OpenSearch
  • Redis
  • React

Outcome

What the build achieves.

Illustrative outcomes from the demonstration environment. They describe the behaviour the design targets — they are not measured client results.

Access control
Source ACLs enforced at retrieval time, before ranking

Engineering outcome of the design. Illustrative, not a measured client result.

Provenance
Every answer assembled only from cited retrieved spans

Architectural constraint.

Refusal behaviour
Declines below a configured support threshold

Design constraint, tuned per deployment.

Regression control
Fixed evaluation set graded on every index, prompt or model change

Method. Accuracy is reported per deployment against that set, never quoted in advance.

Demonstration figures — design intent, not client-verified

Related

Other demonstrations.

Start a conversation

Bring us the version of this that is actually yours.

We will read the constraint, say what we would change, and be direct about what we do not yet know.

Start a ConversationBack to case studies