In this lesson you’ll cover:
- Why “the page is read by humans, the chunk is retrieved by machines” — and the gap between the two
- The three numbers that define RAG: ~500 tokens per chunk, 3–8 chunks per query, and zero surrounding context
- The four-stage RAG pipeline (ingestion/chunking, embedding, retrieval, generation) and the failure mode at each
- Why chunking happens before any query is known, and the rule that any ~400-word window must stand alone
- Modular vs non-modular content — how the same facts extract cleanly or fail, with worked examples
- Why headings are a retrieval signal — a question-heading matches queries; a vague one gives the slot away
- How synthesis picks chunks: self-contained, scoped, sourced chunks win; broken-reference chunks get dropped
- Why attribution must be inline, since bylines, dates and reference lists don’t travel inside a chunk
- The four failure patterns (narrative build-up, distributed definition, table without context, multi-claim paragraph) and their fixes
- The four-part “proposition” structure and a 5-check extraction-readiness audit