About This Skill

Every RAG tutorial shows the same pipeline; almost none of it survives contact with your actual corpus. Naive chunking destroys context across boundaries. Cosine similarity retrieves semantically adjacent text that doesn't answer the question. Re-ranking adds latency without improving precision when the problem is in the chunks, not the ranking. This skill designs your RAG system for your specific data: the chunking strategy that preserves meaning (fixed-size, semantic, document-aware — different corpora need different approaches), the embedding choice and retrieval architecture, the hybrid search setup if your documents have structured metadata, the re-ranking layer if you actually need it, and the evaluation framework so you know when retrieval quality degrades. Give it your corpus characteristics, your query patterns, and the quality bar you need; it returns a RAG architecture calibrated to your problem instead of copied from a tutorial.