rag-engineer · davila7 — Claude Code skill: RAG pipelines and vector search

What rag-engineer does

The skill frames the agent as a RAG systems architect and steers how it works with documents and search: vector embeddings, semantic search, chunking, retrieval pipeline design, hybrid keyword plus semantic search. Inside are patterns and a sharp-edges table with severity and a fix for each row.

The patterns are concrete. Semantic chunking follows sentence boundaries and topic shifts instead of a token limit. Hierarchical retrieval indexes at several chunk sizes, runs a coarse pass for candidates, then a fine-grained pass for precision. Hybrid search pairs BM25 with vector similarity and combines the scores through Reciprocal Rank Fusion.

Anti-patterns get their own section: fixed chunk size, embedding everything, skipping evaluation. The skill tells you to measure retrieval quality separately from generation, because one blended number hides where the pipeline actually breaks. Prerequisites are LLM fundamentals, a grasp of embeddings and basic NLP concepts.