Treating chunking as design rather than preprocessing matches my experience that most "the model is bad" tickets are really retrieval tickets. The silent-truncation trap is the one that burns people: feeding an 800-token chunk to a 512-ceiling embedder drops the tail with no error. Do you tune chunk size per document type, or hold one recursive default across the corpus?
Kartik N V J K
AI Developer | Making AI reliable, trustworthy & accessible to everyone | Active community contributor
Treating chunking as design rather than preprocessing matches my experience that most "the model is bad" tickets are really retrieval tickets. The silent-truncation trap is the one that burns people: feeding an 800-token chunk to a 512-ceiling embedder drops the tail with no error. Do you tune chunk size per document type, or hold one recursive default across the corpus?