A bit of both, but for different geometric reasons...
# 🧶︱general
w
A bit of both, but for different geometric reasons. On long docs, standard chunking breaks up continuous context windows splitting cause from effect across arbitrary token boundaries and causing the model to hallucinate missing connections. On short API refs, it’s a density problem: when a 2-line code example or parameter definition gets sliced in half at a chunk edge, the functional payload loses its structural integrity completely. In our testing, the 7% retrieval recovery came from replacing fixed-token slicing with dynamic boundary checks treating the document as a continuous manifold so related parameters and context stay anchored together regardless of page length.