Retrieval Latency and Performance on NQ
1.96H2D Latency (ms)AgentIR
Evaluation Results
| Method | Links | |||||||
|---|---|---|---|---|---|---|---|---|
| AgentIRHardware=A100 40 GB, Batch size=all judged queries, chunked when B · N · 4B > 16 GB, k=100, |D|=2.7M, |Q|=3,4522026.05 | 1.96 | 7.04 | 284.9 | 0.05 | 0.82 | 3.4 | 99.4 | |
| AgentIRHardware=A100 40 GB, Batch size=all judged queries, chunked when B · N · 4B > 16 GB, k=32, BLOCK=128, |D|=2.7M, |Q|=3,4522026.05 | 1.96 | 7.07 | 79.3 | 0.04 | 0.33 | 8.4 | 99.5 |