Your decoupled-and-aligned index is now 171x larger tha...

To address the storage and performance issues, you must treat them as distinct problems. For storage, you can apply residual or product quantization to the patch embeddings, which can significantly reduce the memory footprint (e.g., 6–10x reduction). For query-time costs, you must ensure that a first-stage pooled-vector retrieval is used to narrow the candidate set before the more expensive late interaction process runs. If this funnel is not in place, the system will be inefficient regardless of the storage strategy.