Explain the execution flow of a RAG pipeline built with LanceDB as the retrieval layer?

A typical retrieval-augmented generation pipeline using LanceDB moves through an ingestion phase (done once, or incrementally as new content arrives) and a query-time phase (done for every user request), with LanceDB serving as the shared retrieval layer connecting the two.

During ingestion, source documents are split into manageable chunks (since embedding models and LLM context windows both have practical size limits), each chunk is embedded — often automatically via LanceDB's embedding function registry — and stored in a table alongside metadata like source document ID, section, or date, which later enables filtered retrieval.

At query time, the user's question is embedded the same way, and LanceDB retrieves the most relevant chunks via vector search (or hybrid search, if exact keyword matching also matters for the domain); those retrieved chunks are then inserted into a prompt alongside the original question, giving the LLM concrete, grounded context to answer from rather than relying solely on its own training data — which is the core mechanism that makes RAG reduce hallucination and allow answers grounded in a specific, up-to-date document set.