Skill Claude CodeCodex
Route each RAG request at runtime to lexical, semantic, or hybrid retrieval, with an optional rewriter for semantic retrieval, optional reranking, and grounded generation. Use when the best retrieval strategy cannot be fixed before execution.
Skill Claude CodeCodex
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 Claude CodeCodex
Use when a RAG request needs bounded multi-round evidence gathering with Critic-based sufficiency checks before returning an answer.
Skill Claude CodeCodex
Arrange parallel RAG with optional LLM query rewriting, two to four parallel retrieval branches, optional per-branch reranking, reciprocal rank fusion, and grounded generation. Use when complementary retrieval routes or multiple query views should be combined.
Skill Claude CodeCodex
Arrange a sequential RAG workflow with one optional query rewriter, one required retriever, one optional reranker, and one required grounded generator. Use after Manage selects a simple one-route retrieval and generation process.
Skill Claude CodeCodex
Rerank candidate documents by cross-encoder relevance score with a local BGE reranker model (BAAI/bge-reranker-large). Use after first-stage retrieval when semantic precision matters more than latency.
Skill Claude CodeCodex
Retrieve and rank title-and-text documents with a field-aware BM25F implementation. Use for exact names, identifiers, quotations, rare terms, and queries with reliable lexical overlap.
Skill Claude CodeCodex
Classify a RAG request into a constrained execution route by calling the frozen Executor Model. Use inside an Agentic workflow before route-specific retrieval or processing.
Skill Claude CodeCodex
Critique a generated answer against the question and supplied evidence by calling the frozen Executor Model. Use for answer quality checks after generation.
Skill Claude CodeCodex
Generate an answer grounded in selected documents by calling the frozen Executor Model through the runtime context. Use as the final generator in Agentic RAG workflows.
Skill Claude CodeCodex
Generate exactly one hypothetical answer-like document from the original query with single-sample HyDE (N=1) for zero-shot semantic retrieval. Use before a vector retriever when query-document wording mismatch is likely; never treat the generated text as factual evidence.
Skill Claude CodeCodex
Retrieve and rank documents by cosine similarity using the embedding service supplied by the runtime context. Use for paraphrases, semantic matching, and weak lexical overlap.
Skill Claude CodeCodex
Analyze a RAG task and produce guidance for choosing exactly one available Agentic RAG Skill using evidence dependencies, retrieval routes, corpus characteristics, and execution budget. Use as the first and only Skill loaded during the Manage selection stage.