RAG explained: answers grounded in your information

Retrieve before generating

Retrieval-augmented generation, or RAG, combines search with a language model. The application finds relevant material, supplies selected passages with the question, and asks the model to answer using that evidence. It can bring organisational information into a response without retraining the model for every document update. RAG is an application pattern, not a guarantee of accuracy.

The information pipeline matters

A typical design prepares documents, divides them into useful passages, indexes them and retrieves relevant results at question time. Poor extraction, outdated files or missing permissions can undermine the answer before generation starts. Retrieval quality and answer quality need separate evaluation: did the system find the right evidence, and did it represent that evidence correctly?

An illustrative policy question

Imagine a team asking which purchasing approval applies to an unusual order. A helpful answer points to the current policy and makes gaps visible. If the policy is ambiguous, the application should route the question to its owner. An indexed document cannot approve a purchase, and an answer must not override Business Central’s validation or authorisation rules.

Questions to ask before investing

Who maintains the source material? How quickly do changes reach search? Can two employees with different access see different evidence? What happens when sources disagree? Start with a small, curated collection and questions with known answers. Frontier AI Works can help teams assess knowledge readiness and evaluate whether retrieval improves a specific task.

Microsoft sources and further reading

• Microsoft Architecture Center: RAG design and evaluation