Hybrid Retrieval Benchmark

Quantization/datatype schemes × prefetch strategies × rescorers, benchmarked against two datasets on Qdrant Cloud with server-side inference.

msmarco jordane95/msmarco-passage-corpus-with-query — 100K passages, real associated search queries dbpedia BeIR/dbpedia-entity-generated-queries — 100K passages, one generated query per passage scifact BeIR/scifact-generated-queries — 15.4K passages, one generated query per passage (100 sampled queries)
Supplementary charts
recall / mrr / ndcg by quantization config, colbert (msmarco) recall / mrr / ndcg by quantization config, colbert (dbpedia) recall / mrr / ndcg by quantization config, colbert (scifact)
Pre-embedded (BEIR / Cohere Embed v3) charts — scidocs
recall / mrr / ndcg by quantization config, colbert p50 & p99 latency by quantization config, colbert throughput (qps) by quantization config, colbert
Pre-embedded (BEIR / Cohere Embed v3) charts — trec-covid
recall / mrr / ndcg by quantization config, colbert p50 & p99 latency by quantization config, colbert throughput (qps) by quantization config, colbert
Pre-embedded (BEIR / Cohere Embed v3) charts — cqadupstack-unix
recall / mrr / ndcg by quantization config, colbert p50 & p99 latency by quantization config, colbert throughput (qps) by quantization config, colbert
Pre-embedded (BEIR / Cohere Embed v3) charts — cqadupstack-gaming
recall / mrr / ndcg by quantization config, colbert p50 & p99 latency by quantization config, colbert throughput (qps) by quantization config, colbert
Pre-embedded (BEIR / Cohere Embed v3) charts — cqadupstack-android
recall / mrr / ndcg by quantization config, colbert p50 & p99 latency by quantization config, colbert throughput (qps) by quantization config, colbert