rag - Retrieve, rerank, ground, cite, and refuse RAG pipeline
Build grounded Q&A pipelines over a corpus by chunking, contextualizing, hybrid retrieval, reranking, grounding, citing chunk IDs, and refusing weak context.
Tags
Updated: 2026-09-24Capabilities
Typical Inputs
Typical Outputs
What this skill does
- Chunk text into structured splits
- Prepend contextual blurbs to chunks
- Retrieve candidates using hybrid search
- Rerank candidates using cross encoders
- Ground answer prompts with context
- Cite chunk IDs in answers
- Refuse queries on weak context
- Evaluate retrieval and answer metrics
Inputs
- Text corpus or extracted documents
- User questions
- Golden Q/A evaluation set
Outputs
- Grounded answers with citations
- Refusal messages for missing context
- Structured chunks with chunk IDs
- Reranked top candidate passages
Requirements
- LLM for contextualization and generation
- Cross-encoder reranking model or API
- Vector database supporting hybrid search
- Clean text input