Learn · How RAG Works
Retrieval-Augmented Generation
RAG lets language models answer questions by retrieving relevant facts from a knowledge base, then generating a grounded response. Ask a question below and watch the retrieve → augment → generate pipeline in action.
1 Retrieve
Score documents by relevance and retrieve the top matches.
Type a query to see retrieved documents.
2 Augment
Inject retrieved documents into the prompt as context.
No query yet.
3 Generate
Compose a grounded answer with citations.
No answer yet.
Why RAG?
LLMs have a knowledge cutoff and can hallucinate. RAG solves both: retrieve up-to-date facts from a trusted knowledge base and ground the model's answer in those sources. Every claim can be traced back to a document, making outputs verifiable and trustworthy.
This demo uses keyword-based retrieval (word overlap, ignoring stopwords). Real RAG systems — including the upcoming Ask bigspark assistant — use embedding-based semantic search to find documents that match the meaning of the query, not just the words. For more on embeddings, see the Embeddings playground.