Skip to main content
Datenschaftler

Solution pattern · reference architecture

Permission-aware enterprise search with citations.

This pattern connects internal document sources with hybrid search and a generated answer. Permissions are applied before retrieval, sources remain visible, and missing evidence does not produce an invented answer.

Starting point

Knowledge is distributed, but central search must not remove access boundaries.

Policies, manuals, contracts, and case files sit in different repositories. A fluent answer is dangerous when document version, source, or permission is unclear.

Business value

Research becomes discoverable, reviewable, and embedded in existing permissions.

Employees receive sources with an answer and can inspect the relevant passage. Business owners see which questions remain uncovered and where document quality or metadata is missing.

What becomes measurable

  • Retrieval quality and source coverage for a versioned set of real questions
  • Permission tests, denied access, and results without sufficient evidence
  • User feedback, escalation reasons, and freshness of indexed sources

Reference architecture

Components and their purpose

Identity, document permissions, search, model, and quality assurance remain separate components. The answer receives only context that the signed-in user is allowed to see.

  1. 01

    Capture sources and permissions

    Approved repositories provide content, version, owner, and access metadata. Microsoft Entra ID supplies user and group context.

  2. 02

    Prepare content

    Azure Functions and, where needed, Azure AI Document Intelligence extract text and structure. Sections remain linked to the document and location.

  3. 03

    Build the search index

    Azure AI Search stores text, vectors, metadata, and security filters for hybrid search. Updates and deletions follow the source.

  4. 04

    Filter access before retrieval

    The application translates Entra context into documented filters. Unauthorised content reaches neither search results nor model context.

  5. 05

    Generate a cited answer

    Azure OpenAI Service formulates an answer from retrieved passages and cites its sources. When evidence is weak, the application shows a clear limit.

  6. 06

    Test quality and misuse

    Application Insights, versioned test questions, content filters, and business feedback support operations, security testing, and model changes.

Technology

Concrete services for implementation

The selection is adapted to existing contracts, regions, security requirements, and the actual scope.

  • Azure AI Search
  • Azure OpenAI Service
  • Microsoft Entra ID
  • Azure Blob Storage
  • Azure AI Document Intelligence
  • Azure Functions
  • Azure Key Vault
  • Application Insights

First project scope

A pilot needs clear boundaries

The first deployment tests data, integration, and the working process in a limited area. It is not a premature enterprise rollout.

Deliberately included

A bounded knowledge domain, a limited approved document set, existing Entra groups, real test questions, and an interface with citations, feedback, and escalation.

Deliberately excluded

No enterprise-wide search index, open internet sources, autonomous write actions in business systems, employment or legal decisions, or replacement of authoritative original documents.

Prerequisites and constraints

Technology does not replace data accountability

Data access, responsibilities, licences, and operations must be clear before implementation. Open points are treated as project risks.

  1. 01Documents need valid versions, owners, metadata, and reliable access rights. Unresolved broad permissions are a security risk.
  2. 02Specialists must define real questions, expected sources, and unacceptable answers as a test set.
  3. 03Regional Azure OpenAI availability, model quotas, content filters, licences, and ongoing search and model costs must be checked in advance.
  4. 04Knowledge owners, identity and security operations, data protection, the business function, and where relevant the works council must be involved.

Germany and the EU

Compliance follows the specific purpose

The index and logs can contain personal or confidential information. GDPR requirements for purpose, minimisation, deletion, data-subject rights, and access also apply to question logs and feedback. EU regions and the processing paths of search and model services must be checked for the selected deployment. The intended purpose, transparency, human oversight, and required documentation must be classified under the EU AI Act. If questions or usage data can support employee monitoring, the works council must be involved early under the Works Constitution Act.

Free initial assessment

Which knowledge question needs both a source and an access check today?

The discovery call frames the document set, permissions, test questions, and the assistant's boundary.

The initial consultation and joint use-case discovery are free and non-binding.