agentic-hybrid-rag

Classify each request as non-retrieval, single-step, or multi-step, then execute the smallest suitable grounded RAG workflow with optional rewriting, reranking, and critique.

SKILL.md

GitHub

Hybrid Agentic RAG

Use the fixed classifier in complexity mode before execution.

Slots

  • classifier: select component-classifier.
  • retriever: select one retriever.
  • rewriter: optional retrieval-query rewriter.
  • reranker: optional evidence reranker.
  • generator: select grounded generator.
  • critic: select answer critic for multi-step execution.

Routes

  • non-retrieval: call the Generator with the original query and no documents.
  • single-step: optionally rewrite the query, retrieve once, optionally rerank, then generate.
  • multi-step: repeatedly retrieve, accumulate new documents, optionally rerank, generate, and critique until approved, no new evidence, or max_iterations is reached.

Always use the original query for generation, reranking, and critique. Rewritten queries are retrieval aids only. Classifier output and generated rewrites are never evidence.

Run scripts/workflow.py:run(request, components) after all required slots are bound.