Users type exact part numbers and our dense retriever m...

Increasing dimensions (e.g., from 128 to 768) simply multiplies index bytes without changing how the model allocates coordinates to distinctions. Tokens that appear rarely in a large corpus will not earn meaningful representation regardless of dimension size. A better approach is to use rank-based fusion to combine dense results with a sparse retriever like BM25, which handles exact matches effectively without requiring complex score calibration.