https://sbeconis.com/writing/ Technical writing by Saulius Beconis. en https://sbeconis.com/writing/evaluating-rag-without-fooling-yourself/ https://sbeconis.com/writing/evaluating-rag-without-fooling-yourself/ How to evaluate RAG honestly — a test set from real questions, retrieval measured separately from generation, and the usual self-deceptions avoided. Wed, 07 Oct 2026 00:00:00 GMT https://sbeconis.com/writing/hybrid-search-and-reranking-measured/ https://sbeconis.com/writing/hybrid-search-and-reranking-measured/ BM25, dense, hybrid and reranked retrieval measured on three BEIR datasets. When the textbook RAG pipeline pays for its latency and when a single embedding model wins. Wed, 07 Oct 2026 00:00:00 GMT https://sbeconis.com/writing/validating-ai-document-extraction/ https://sbeconis.com/writing/validating-ai-document-extraction/ Measured on 200 real receipts — how plain-code validation decides which AI-extracted documents can skip human review, and what it still misses. Wed, 07 Oct 2026 00:00:00 GMT https://sbeconis.com/writing/when-not-to-use-an-llm/ https://sbeconis.com/writing/when-not-to-use-an-llm/ A practical decision test for engineers and buyers — when a language model is the right component, and when code, search or a human is cheaper and more reliable. Wed, 07 Oct 2026 00:00:00 GMT https://sbeconis.com/writing/why-rag-retrieval-fails/ https://sbeconis.com/writing/why-rag-retrieval-fails/ Most bad RAG answers start with bad retrieval. A field guide to the failure modes — parsing, chunking, vocabulary mismatch, ranking depth — and how to diagnose them. Wed, 07 Oct 2026 00:00:00 GMT