ai-rag

RAG retrieval systems combine chunking strategies, hybrid search (BM25 + vector), and reranking to retrieve relevant context. This Skill designs end-to-end pipelines with semantic chunking, cross-encoder reranking, and query rewriting to ground generation on accurate, ranked results at scale.

Hybrid retrieval combines BM25 lexical matching with vector similarity search, capturing both exact keyword matches and semantic relevance. Vector-only search misses keyword-exact results; hybrid retrieval with reranking improves recall@K and ranking quality for production systems.

Retrieval evaluation uses metrics like recall@K, nDCG, and MRR to measure ranking accuracy and groundedness. This Skill provides evaluation components to validate that your retrieval pipeline returns relevant, grounded context for generation.

Yes. Cross-encoder reranking re-scores retrieved results for relevance; query rewriting reformulates user queries to improve matching. Together with adaptive retrieval patterns, these techniques increase ranking precision and recall in large-scale knowledge bases.

Page-level and semantic chunking partition documents into retrievable units that preserve context. Semantic chunking groups related content; page-level maintains boundaries. This Skill supports both to ensure accurate grounding and minimal context loss in RAG applications.

Context injection embeds retrieved results into prompts; prompt caching stores repeated context to reduce latency and cost. This Skill packages both patterns with grounded generation to enable production RAG workflows at scale.