← InsightsAI knowledge & enterprise search

Enterprise AI search and internal knowledge bases: accurate, permission-aware answers

The value of an internal AI search system is not that it can produce fluent answers. The hard part is retrieving the right evidence, showing only content the user is allowed to see, linking back to sources and refusing to guess when evidence is weak.

01

1. Make source systems consistently searchable first

SharePoint, OneDrive, internal wikis, file stores and databases all have different metadata, permissions and update behaviour. A useful RAG system needs a traceable index that keeps those differences under control.

  • source URL and document ID
  • version and modified time
  • business metadata
  • deletion and refresh synchronisation
02

2. Vector-only retrieval is rarely enough

Semantic similarity helps conceptual questions, but business users also search for product codes, contract IDs, names, dates and exact terms. Hybrid keyword and vector retrieval followed by reranking often works better.

  • full-text / BM25
  • vector search
  • hybrid retrieval
  • semantic reranking
03

3. Enforce permissions before content reaches the model

The retrieval layer should filter documents using the caller's real permissions before any content is sent to the language model.

  • Entra identity
  • group and document-level permissions
  • security trimming
  • least-privilege service identity
04

4. Chunking, metadata and freshness work together

Chunks that are too large add noise; chunks that are too small lose context. Titles, sections, document type, dates and other metadata help retrieval filter and rank more precisely.

  • meaningful section boundaries
  • metadata filters
  • prefer current content
  • exclude archived versions
05

5. Answers must be verifiable

Enterprise answers should carry source links. If retrieval does not find strong evidence, the system should surface uncertainty instead of inventing a plausible answer.

  • source links
  • document title and date
  • controlled no-answer behaviour
  • grounded uncertainty
06

6. Without an evaluation set, improvement is guesswork

Create a test set from real business questions and measure retrieval quality separately from final-answer quality.

  • top-k retrieval hit rate
  • grounded-answer checks
  • permission tests
  • freshness and no-answer tests
Related service

AI for business

Document and email processing, information extraction, enterprise search, assistants and agentic workflows.

View the service →
First step

Show us the process you would like to make simpler.

In a 30-minute call we look at how it works today, what is worth automating, where AI can help, and what should stay simple.