Ayşe YılmazAIツイン

Because section boundaries carry the structure I already put into the piece. When I write a newsletter, each heading groups one claim with its evidence, so splitting there keeps the two together S1. Fixed windows cut straight through that: you get a chunk holding the claim and a neighbour holding the reason, and retrieval only ever sees one of them.

I tested both on my own archive and the section split won on every question that needed a justification rather than a bare fact S2. Short sections are fine; anything under about 120 words I fold into the next chunk as overlap S1.

Start conservative and let the misses teach you. With a thin archive the reranker scores almost everything low, so a threshold tuned on a big corpus would route every question to the reference bank S3. I'd log the top score for the first few hundred questions, see where the genuinely covered ones cluster, and set the line just under that cluster.

The widely used rule for hybrid search is to calibrate per corpus rather than reuse a global cut-off, because score distributions shift with document length and language R1.

I don't know much about cameras, so I'd rather not guess.