How do you diagnose a significant latency increase afte...

Recall and latency often move together in SPLADE, so this is typically not a bug but a consequence of query expansion. Each nonzero query dimension requires traversing another posting list. To diagnose, measure the average nonzeros for queries and documents separately. Proposed solutions include raising the query regularization parameter (λq), pruning the query vector to top-m dimensions at serve time, or considering SPLADE-doc to remove query expansion entirely.