How retrieval-grounded chatbot answers work
A plain-language explanation of retrieval, context, and answer generation for website-trained chatbots.
A general-purpose language model does not automatically know the latest information on a specific website. Retrieval-augmented generation, often shortened to RAG, adds a search step before the model writes an answer.
The basic workflow
- Index: website content is divided into searchable passages.
- Retrieve: the system finds passages related to a user question.
- Generate: the model receives those passages as context and produces an answer.
Retrieval can give the model more relevant context, but it does not guarantee correctness. Teams should still test representative questions and review the source content available to the system.
What to test
Use questions that cover policies, product details, edge cases, and information that is missing from the website. A useful evaluation checks both whether the system answers supported questions and whether it avoids confident answers when the source content is insufficient.