- You cannot control the chunks. OpenAI File Search cuts at 800 tokens with 400 overlap; other systems cut at 256 or 512 tokens, sentence boundaries, or character counts; web pipelines use whatever the snippet extractor produced.
- Chunk boundaries are rarely fatal — 50% overlap means a good answer sentence cut in one window survives intact in the next.
- Rewriting every paragraph as a self-contained answer machine produces robotic text humans stop reading — and humans still assign the links and mentions that get you retrieved at all.
- Local self-containedness: any few-hundred-word window of the page should make sense alone — name things explicitly instead of “it/this tool”, keep the answer near its subject.
- One direct-answer sentence per target question — the highest-leverage edit measured; see Answer shape for AI citations.
