Structure-aware and contextual chunking, measured

The documents of three near-identical matters go into one index with no matter filter, so a chunk is only found if it carries which property it belongs to. That is the failure contextual headers and generated context exist to fix. The contextual pass runs on your Claude subscription through claude -p, with the whole document as the system prompt so that it is the same cacheable prefix for every chunk of that document; embeddings go through the OpenAI SDK. The decision that matters most is how you score: recall@k on its own rewards big chunks, so the table also reports recall when every strategy may hand the model the same 2,000 tokens.