RAG and Knowledge Engineering

RAG and Knowledge Engineering

Hybrid Search — References

Hybrid search › References

36 of 81 in this book · 2 min read

The two retrieval families being combined, and the fusion rules that combine them.

The lexical arm

  • The Probabilistic Relevance Framework: BM25 and Beyond — Stephen Robertson and Hugo Zaragoza, Foundations and Trends in Information Retrieval, 2009. The derivation and the parameters of the term-frequency ranking that is still the default lexical arm. Cited by name.
  • SPLADE: Sparse Lexical and Expansion Model for First Stage Ranking — Thibault Formal, Benjamin Piwowarski and Stéphane Clinchant, ACM SIGIR, 2021. Learned sparse representations that keep term-level matching while expanding vocabulary, which is the modern alternative to a classical analyser. Cited by name.

The dense arm

  • Dense Passage Retrieval for Open-Domain Question Answering (Karpukhin et al., 2020), arXiv:2004.04906 — the dual-encoder construction and the comparison against lexical baselines that established where each one wins.
  • ColBERT: Efficient and Effective Passage Search via Contextualized Late Interaction over BERT (Khattab & Zaharia, 2020), arXiv:2004.12832 — late interaction, a third retrieval family that behaves unlike either arm discussed here.

Combining rankings

  • Reciprocal Rank Fusion Outperforms Condorcet and Individual Rank Learning Methods — Gordon V. Cormack, Charles L. A. Clarke and Stefan Buettcher, ACM SIGIR, 2009. A parameter-light rank fusion rule, and the evidence that it beats more elaborate score-based combinations. Cited by name.
  • Combination of Multiple Searches — Edward A. Fox and Joseph A. Shaw, TREC-2 proceedings, 1994. The original score-combination rules and the normalisation problem they expose. Cited by name.

Where the two families actually differ

  • BEIR: A Heterogeneous Benchmark for Zero-shot Evaluation of Information Retrieval Models (Thakur et al., 2021), arXiv:2104.08663 — per-dataset results showing a strong lexical baseline beating dense retrieval on several domains, which is the empirical case for running both.

No verified reference is recorded for per-arm depth selection, for query-shape routing between arms, or for offset-based deduplication across differing segmentations. These are engineering practice rather than published results.