Part 1 of 5 · 2 min read
Overview
In this module you will
- Explain why multi-hop questions defeat a single search, and show it with numbers.
- Describe agentic RAG: a loop in which the agent plans, searches, checks what it found, and decides what to do next.
- Describe GraphRAG: extract entities and relationships into a graph, then answer by walking it.
- Compare single-pass, agentic and graph retrieval on accuracy, cost, freshness and failure modes.
- Say honestly when the extra machinery is not worth it.
Plain RAG is "search once, answer". That is fine for "What does E-4021 mean?", but real questions chain: Which office does the manager of the team that owns the billing service work in? No passage contains that sentence. The answer is spread over four short facts, and the question shares words with only the first. An agent is exactly the kind of system that can keep going: use what it found to decide what to ask next. GraphRAG attacks the same problem differently, by connecting the facts ahead of time so a question becomes a walk along edges. Both cost more than a single search, so you should know when they earn it.
About the examples
The corpus is an invented 20-passage company wiki: 13 one-fact passages (who owns which service, who manages which team, where people sit, when offices opened) plus 7 distractors that reuse the same words. The 24 questions are mine: 6 one-hop, 6 two-hop, 4 three-hop, 3 four-hop, 2 list questions and 3 with no answer. The "language model" parts are rule programs: a planner that knows 11 question shapes, a reader that extracts facts with a few text patterns, and a query writer that fills in templates. They show the mechanism; a real model reads any phrasing and also makes mistakes the examples do not. The tiny, clean corpus makes the agentic loop look nearly perfect, which real systems will not be. Read the numbers as direction, not as benchmarks. No real model or service was called.