← All questions
Inference & ServingVery Hard· PLAID· Latency· Retrieval Evaluation

A teammate wants to remove PLAID’s centroid-interaction...

The debate over removing the centroid-interaction stage should be settled by empirical measurement rather than theoretical argument. You should compare the recall of the centroid-interaction stage against exhaustive MaxSim over the same candidate set. If the stage is highly accurate, it provides a massive latency benefit for free. If it causes a measurable drop in recall, the decision becomes a product trade-off regarding whether the latency gain justifies the loss in retrieval quality.

Code playground (Python)