rag-systems
Installation
SKILL.md
rag-systems
Core Philosophy
Retrieval-Augmented Generation (RAG) is not simply splitting a PDF into 500-character chunks, embedding them with OpenAI, and dumping them into a vector database. Naive RAG produces fragmented context, keyword mismatch, and hallucinated answers. Production-grade RAG is a precision information retrieval pipeline that combines semantic chunking, hybrid search (dense embeddings + sparse BM25), cross-encoder reranking, and rigorous automated evaluation using the RAG Triad.
4-Step Production RAG Engineering Pipeline
Step 1: Chunking Strategy & Document Decomposition
- Avoid Arbitrary Fixed-Length Splitting:
- Fixed character counts cut mid-sentence, slicing critical context in half.
- High-Precision Chunking Archetypes:
- Semantic / Markdown Chunking: Split strictly along document boundaries (
#,##,###, code blocks). - Recursive Character Chunking with Overlap: Split by paragraph (`
- Semantic / Markdown Chunking: Split strictly along document boundaries (
), then sentence (
`), with 15–20% token overlap to preserve transitional context.
- Parent-Document / Small-to-Big Retrieval: Embed small chunks (128 tokens) for precise vector matching, but return the larger parent chunk (1,024 tokens) to the LLM context for generation.