multi-hop RAG

Some questions cannot be answered by a single lookup. "Which university did the author of this book attend?" needs two hops: first find who the author is, then find where they studied — and the second search depends on the answer to the first. Multi-hop RAG retrieves in steps, using what it has found so far to shape the next search.

Instead of one retrieve-then-answer pass, the system loops: retrieve, read, decide what is still missing, form a new query from that gap, retrieve again, and continue until it has enough to answer. This handles compositional and comparative questions that single-shot retrieval simply fails on, but it costs more model calls and more latency, and it can drift off course if an early hop fetches the wrong thing and every later hop builds on that mistake. It shades naturally into agentic RAG.