EAHR hybrid retrieval matches exact results at 30x lower latency

Modern retrieval-augmented generation (RAG) systems typically fuse fixed Top-L results from dense and sparse retrievers, treating anything beyond that cutoff as zero. But truncated fusion isn't always equivalent to complete-list fusion—unread cross-list ranks can change Top-K membership or ordering. Chunran Zhang's arXiv paper (2608.07152) introduces Exact Adaptive Hybrid Retrieval (EAHR), which targets the exact ordered Top-K defined by complete-list weighted RRF (Reciprocal Rank Fusion) and treats channel depth as request-specific state, not a preset constant.

EAHR leverages two scalable mechanisms: Per-Vector Scalar Quantization (PVS) for densely encoded vectors and Posting Block-Max (PBM) for sparse inverted indexes. Both produce resumable rank lists, meaning the system can pause and request more candidates from a channel without restarting. Fusion bounds estimate the maximum impact of unread contributions, and EAHR only requests additional ranks when they could actually change the Top-K output. Every successful request matches complete-list fusion exactly; otherwise, execution safely continues to list exhaustion.

Across five test collections and five temporal corpus snapshots, fixed depths tuned on historical queries failed to transfer reliably, while complete-list weighted RRF remained competitive. EAHR reproduced the complete-list ordered Top-20 in all 150 query-snapshot combinations. Under a warm-cache, interleaved, order-balanced protocol, the paired geometric-mean latency ratios of exhaustive batch execution to EAHR were 23.35 on TREC-DL 2019 and 30.28 on TREC-DL 2020.

EAHR doesn't guarantee a speedup for every request—anti-correlated rankings exhausted both lists, and some difficult queries ran slower. But it fixes the exact result while adapting execution depth to current rankings, offering a principled alternative to arbitrary Top-L cutoffs for production RAG pipelines.