CHAPTER 03 · The Taxonomy of Agentic RAG Systems · 9 / 9
Key takeaways
- Single-agent router: one agent routes each query to the best source. Simple, efficient, best for limited toolsets.
- Multi-agent: specialized agents retrieve in parallel under a coordinator. Scalable and modular, but harder to coordinate.
- Hierarchical: tiered agents with strategic top-level prioritization. Good for complex queries, more orchestration overhead.
- Corrective: a five-agent loop that evaluates and fixes its own retrieval. High factuality, added latency.
- Adaptive: a classifier matches retrieval effort to query complexity, trading some correctness for responsiveness.
- Graph-based (Agent-G): combines knowledge graphs and text with a critic and feedback loops; strong multi-hop reasoning.
- Agentic document workflows: stateful, end-to-end automation of document processes, beyond single-question RAG.
- The taxonomy is a ladder of increasing autonomy and structure, each rung adding capability at the cost of complexity.