Search

What are you looking for?

Search our services, use cases and practical insights.

Enter at least 2 characters

Popular starting points

Knowledge, data and RAG

How is an AI knowledge base maintained and evaluated over time?

A knowledge base is not finished after its first import. Sources, permissions, questions and models change, so RAG needs both a domain and technical operating process.

The short answer

An AI knowledge base stays reliable when source ownership, synchronisation, permissions, quality measurement, monitoring and error handling are organised as recurring operational responsibilities with clear owners.

In brief

  • Business teams own content and validity; technical teams own processing and availability.
  • Every source needs a defined update, failure and deletion process.
  • Retrieval and answer quality are monitored with stable test questions and real feedback.
  • Changes to models, embeddings, chunking or sources must be versioned and tested before release.

Which roles does operation require?

Technology alone cannot decide which statement is authoritative. Domain owners approve sources and resolve conflicts; platform and operations teams secure data flow, permissions and availability. A named product owner connects value, quality and change.

Which roles does operation require?
Operational taskLead responsibilityCadence or triggerEvidence
Approve source and validityDomain ownerBefore onboarding and after content change or expiryApproval status, version, scope and accountable person
Operate connector, index and deletionsPlatform or operations teamContinuously; verify after every synchronisation runCount reconciliation, delay, failures and deletion confirmation
Test permissionsEngineering with securityAfter role changes, at release and regularly as regression testsPositive and explicitly denied cases for each role
Assess retrieval and answer qualityProduct owner with the domain teamBefore release and on a fixed monthly or quarterly cycleVersioned test set, results and accepted error limits
Handle false or missing answersSupport route with the domain ownerAfter a report or recurring failure patternClassified case, source correction and new regression test
Handle a security or privacy incidentIncident lead with security and privacyImmediately after detectionImpact, containment, cause, notification and retest

What does the ongoing data cycle look like?

Each source passes through detection, processing, indexing and quality checks. Operations monitor not only technical availability but also delay, expired content and complete removal of deleted documents.

  1. Step 1

    Inventory

    Document the source, owner, purpose, audience and sensitivity.

  2. Step 2

    Synchronise

    Carry text, metadata, permissions and deletions into the index.

  3. Step 3

    Validate

    Run samples, count reconciliation and known test questions.

  4. Step 4

    Monitor

    Watch failures, delay, no-result queries and quality indicators.

  5. Step 5

    Improve

    Feed user feedback and new failure patterns back into sources and tests.

Which quality should be measured?

A good model answer can be based on the wrong result; a poor answer can occur despite the right source. Retrieval and generation should therefore be measured separately and together. Averages must not hide critical failure types.

  • Recall: is the required passage among the returned results?
  • Precision: how many returned passages are actually relevant?
  • Authorisation: is unauthorised content always excluded?
  • Grounding: does the source support the formulated statement?
  • Operations: freshness, latency, cost and failure rate

How are changes and incidents handled?

New models, embedding versions or chunking rules can change results even when source documents stay the same. Such changes need comparison on a fixed test set and a rollback path. Unauthorised or dangerously false answers require a prepared incident process.

  • Version configuration and data state
  • Test changes in a separate environment first
  • Use critical regressions as release blockers
  • Be able to disable affected sources or functions quickly
  • Document cause, impact, correction and retesting
Example from day-to-day business

Example: monthly operation of a service knowledge base

The service team owns manuals and approvals. A daily job carries changes and reports failures. Weekly review covers frequent no-result queries and negative feedback; a fixed test set runs monthly across products, languages and roles. Before a model or embedding change, the team compares old and new and releases only without critical regressions.

What to remember

Plan domain curation, technical synchronisation and measurable quality before the first production use. A knowledge base remains only as good as its operating process.

Sources and further reading

These primary sources provide further detail on definitions, technical foundations or responsible use.

Content reviewed

Would you like to apply this to your situation?

Together, we clarify what makes sense for your process, data and systems – in plain language and without unnecessary complexity.

Discuss Your Project