Agentic Conditional RAG

Conditional RAG Agentic Skill

Route each request at runtime through exactly one of three retrieval strategies: lexical, semantic, or hybrid.

Slot Binding

Bind the slots as follows:

• classifier: select component-classifier.
• rewriter: optionally select component-hyde-rewriter.
• lexical_retriever: select component-bm25-retriever.
• semantic_retriever: select component-vector-retriever.
• reranker: optionally select component-bge-reranker.
• generator: select component-grounded-generator.

Do not exchange the lexical and semantic Retriever bindings.

Routing

Call classifier first with the original query and routing constraints.

• lexical: do not call the Rewriter. Send the original query only to lexical_retriever.
• semantic: if a Rewriter is bound, send its non-empty rewritten_query only to semantic_retriever; otherwise send the original query.
• hybrid: send the original query to lexical_retriever; send the optional rewritten query, or the original query when no Rewriter is bound, to semantic_retriever; fuse the two ranked lists with reciprocal rank fusion.

Reject any classifier route other than lexical, semantic, or hybrid.

Evidence Safety

The Classifier result and Rewriter output control routing and retrieval only. They are never evidence documents.

Always use the original query for optional reranking and final generation. Pass only real documents returned by the Retriever slots to the Reranker and Generator.

Execution

Run scripts/workflow.py:run(request, components) after every required slot is bound. Record classification, rewriting, retrieval, fusion, reranking, and generation decisions in the returned trace.