This path is for you if you want answers from your own files, not from the model's memory. It covers cutting documents into pieces, embeddings, filters, a mix of search types, and a citation you can check. No part is live on this page yet. Start with the Practical AI part on retrieval in plain English, then come back for the build.
- 1 RAG in One Picture: What Retrieval-Augmented Generation Really Does RAG separates knowledge lookup from model reasoning. Embed, retrieve, augment the prompt, then generate. Bad retrieval looks like a broken search engine with better prose. Scheduled · November 24, 2026
- 2 Document Chunking for RAG: Too Big, Too Small, and Just Usable Choose chunking for your docs: fixed, recursive, or semantic, with overlap. Special-case tables and code. Prefer small chunks for search and larger parents for answers when you can. Scheduled · November 25, 2026
- 3 Embedding Models for RAG: Picking Dense, Sparse, and Hybrid Vectors Without Religion Use dense vectors for semantic match, sparse for exact terms, and hybrid fusion when both matter. Consider Matryoshka dims to cut storage. Test on your queries. Scheduled · November 26, 2026
- 4 Metadata Filtering in RAG: The Secret to Security, Speed, and Zero Stale Data Attach tenant, role, freshness, and doc-type metadata to every chunk. Prefer pre-retrieval filters for security. Treat ACL mistakes as incidents, not UX bugs. Scheduled · November 27, 2026
- 5 Citations and Show Me the Source UX: Building Verifiable RAG Applications Ship RAG with visible sources, passage-level cites when you can, and automated checks that quotes exist in retrieved text. Latency trades are real; inventing footnotes is not allowed. Scheduled · November 28, 2026
- 6 RAG Failure Modes: Stale Docs, Wrong Neighbors, and Confident Nonsense Watch for stale docs, bad neighbors, and unsupported claims. Add freshness signals, rerankers, and refusal paths. Treat RAG outages like search outages. Scheduled · November 29, 2026
- 7 Light RAG Evals Non-Engineers Can Run: From Guesswork to Golden Sets Run light RAG evals with a golden question set, simple faithfulness and relevance scores, and a repeatable schedule. Vibe checks do not replace labeled cases. Scheduled · November 30, 2026
- 8 When Fine-Tuning or a Bigger Context Window Is Better Than RAG Choose RAG for changing private knowledge with citations, fine-tuning for tone and rigid formats, and long context for whole-document comparison. Most production stacks blend the three on purpose. Scheduled · December 1, 2026