Every file in the ai-writings corpus, chunked and embedded (bge-m3, 1024-dim, cosine). Queries are embedded the same way and matched by meaning, not spelling. Best chunk speaks for its file.