Two specializations that solve different gaps
Lessons 11 and 12 each fixed a different weakness of plain semantic search: hybrid RAG fixed poor exact-term matching, graph RAG fixed the inability to reason across explicit relationships. They are not competing solutions to the same problem, and a system can legitimately use both, but it helps to be precise about which gap each one actually closes, because reaching for the wrong one wastes real engineering effort.
The diagnostic question to ask
Ask: does this question fail because the right words are not being matched, or because the right connection is not being made? "Show me the warranty terms for model X-4471" is a matching problem, the model number needs to be found exactly, which points at hybrid RAG. "Which of our suppliers are also suppliers to our biggest competitor" is a connection problem, no single document states this, it has to be assembled by walking relationships between supplier and competitor entities, which points at graph RAG.
| Hybrid RAG | Graph RAG | |
|---|---|---|
| Fixes | Poor exact-term matching in pure semantic search | Inability to reason across explicit relationships |
| Setup cost | Low: add a keyword index alongside the vector index | High: entity/relationship extraction and schema design |
| Best data shape | Documents rich in codes, names, precise terms | Domains where connections between entities are central |
| Answers questions like | "Find the exact clause about X" | "Which things are connected to which other things, and how" |
In practice, sophisticated production systems run both, wired together through the modular architecture from Lesson 8: a router sends precision-lookup questions to a hybrid-search module and connection-reasoning questions to a graph module, rather than forcing every question through just one of them.
Checkpoint
- Hybrid RAG fixes a matching problem (the right words not being found); graph RAG fixes a connection problem (the right relationship not being traversed).
- The diagnostic question: does this specific question fail on exact terms, or on needing to connect facts across entities?
- Production systems often run both, routed by question type through a modular architecture, rather than choosing only one globally.
If anything here still feels unclear, ask before moving to Lesson 14.