When an employee asks ChatGPT "How many days of annual leave do I get?", the answer is general labor law, not your company's policy, because the model has never seen your internal documents.
How RAG works
RAG stands for Retrieval-Augmented Generation. Put simply: search first, then answer. When a question comes in, the system first searches your company's document library for relevant material, then passes it to the model along with the question. The model answers from real information and can say which document the answer came from.
Why not train your own model?
- Training is expensive and slow, and every document change means retraining. With RAG, updated documents are usable immediately.
- RAG can cite its sources, so answers can be verified and made-up answers are reduced.
- You can control who is allowed to search which documents, which a trained model cannot do.
Who benefits most
- Organizations with many manuals, SOPs or contracts that staff search often
- Companies with staff turnover that keep repeating onboarding
- Support teams answering the same questions from the same documents
What to watch out for
RAG quality depends on document quality. Outdated or contradictory documents lead to confusing answers, so the first step in every project is to select and organize the documents. Choose a model provider that does not train on your data, or deploy on your own servers for sensitive information.
Learn more about our Internal AI Knowledge Base service.