| name | agentic-rag-router |
| description | Choose the cheapest sufficient retrieval strategy for a query before answering. Use whenever the system has multiple retrievers (vector, hybrid, graph, metadata, summary) and needs to decide how to fetch evidence for a question. Trigger on any retrieval/QA task where query types are mixed. |
agentic-rag-router
The point of agentic RAG isn't "more agents" — it's adaptive retrieval: use the cheapest
sufficient behavior for the query at hand. This skill is the router.
Routing rules
- exact names / IDs / error codes / config keys / quoted strings →
hybrid_search
- simple conceptual / single-hop lookup →
vector_search
- relationship / dependency / impact / lineage / "connected to" →
graph_search
- scoped by region / tenant / version / date / department →
metadata_filtered_search
- broad summary / theme →
hierarchical_summary_search
- ambiguous wording → multi-query expansion
- weak evidence after retrieval → corrective retry (regrade, switch retriever, decompose)
- high-risk answer → add a critic/verification pass
Output (JSON, user-safe — no hidden chain-of-thought)
{"query_type": "...", "selected_tool": "...", "tool_args": {}, "decision_reason": "brief", "needs_fallback": false}
Notes
- Don't answer the user here; only return the routing decision.
- This is the runtime cousin of the comparison harness in
evals/python/rag_compare/, which
measures which strategy actually wins per query class at what cost/latency.
