rag

RAG: make enterprise knowledge safely usable by AI

RAG connects a request with relevant approved knowledge passages and adds source context.

At a glance

The essentials

  1. source selection matters more than volume
  2. permissions need to be preserved
  3. knowledge ownership remains human
Contents

Retrieve first, then work with context

RAG finds relevant passages in a curated collection and provides them as working context. This can ground answers and agent work more firmly in enterprise knowledge.

A knowledge system needs owners

Sources require approval, validity, access rules and accountable business teams. Missing or conflicting evidence must be visible rather than masked by false certainty.

  • curated sources
  • role-based access
  • source context
  • maintenance process

Prerequisites start before technical retrieval

Knowledge domains, document types, owners, sensitivity and validity rules are described first. Only content that is necessary for the purpose and approved should be reachable.

Metadata for source, version, validity and permission helps classify findings correctly and remove outdated knowledge.

  • defined purpose
  • accountable source owners
  • clean and discoverable documents
  • preserved access rules
  • defined maintenance cycles

RAG improves context but does not guarantee a correct answer

Errors may arise from incomplete sources, unsuitable chunking, imprecise queries, wrong retrieval or subsequent generation. Quality and coverage therefore need representative questions, expected sources and difficult counter-examples.

A dependable response should expose its evidence and leave missing or conflicting support visible. Consequential statements still require expert review.

Permissions and knowledge integrity are security concerns

A RAG system must not expose documents that the requesting role is not allowed to access. Its ingestion process also needs to stop unreviewed or manipulated content from becoming an unnoticed basis for work.

BSI identifies manipulated knowledge collections as a possible knowledge-poisoning risk. Approval, provenance, changes and removal of sources therefore need a controlled process.

When RAG is useful and when it is not

RAG suits distributed, changing knowledge where concrete evidence is needed and owners can maintain the collection. A structured database or conventional search may be simpler for a small set of stable rules.

Without adequate source coverage, access design or expert evaluation, a RAG system should not produce binding outcomes.

Source note

This description follows the NIST definition of RAG and risk guidance from NIST and BSI. Data protection requirements need assessment against the actual information estate; this is not legal advice. Primary sources reviewed on 11 August 2026.

Primary sources

Official sources, editorially checked on 11 August 2026.

Which process should work better in your organisation?

We begin with the work, the people and the current process — then assess which form of AI genuinely makes sense.

Discuss your process