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

A retrieval-augmented generation pipeline using Qdrant separates cleanly into an ingestion phase, run once or incrementally as new content arrives, and a query-time phase, run for every user request, with Qdrant's collection serving as the shared retrieval index connecting the two.

During ingestion, documents are chunked, embedded, and upserted into a Qdrant collection with payload metadata like source, date, or category, alongside payload indexes for any fields the application expects to filter on later, such as restricting retrieval to a specific tenant or document category.

At query time, the user's question is embedded the same way, and Qdrant's query_points retrieves the most relevant chunks — combining vector similarity with any applicable payload filters, and optionally fusing dense and sparse retrieval via Prefetch and a fusion method for hybrid search — before those retrieved chunks are inserted into a prompt as grounding context for the LLM to generate its final, evidence-based answer.