---
title: "Hybrid retrieval (IT)"
source: "https://systems-analysis.info/int/Hybrid_retrieval_(IT)"
wiki: "systems-analysis.info/int"
article: "Hybrid_retrieval_(IT)"
language: "it"
categories:
  - "Category:Italian"
  - "Category:Large language models"
  - "Category:Prompt engineering"
revision_id: 3090
wiki_created_at: 2026-09-06T23:15:33Z
wiki_modified_at: 2026-09-06T23:15:33Z
downloaded_at: 2026-09-07T22:55:01Z
---

# Hybrid retrieval (IT)

**Hybrid Retrieval (recupero ibrido)** — classe di metodi di recupero dell'informazione che combinano segnali lessicali (sparse) e semantici (dense/late‑interaction) per aumentare il richiamo e la precisione dei risultati. Gli schemi ibridi uniscono i vantaggi della corrispondenza esatta dei termini (BM25/TF-IDF) e della similarità vettoriale (bi-encoder, modelli multi-vettore a interazione tardiva), e utilizzano metodi di fusione dei ranking resistenti alle differenze di scala degli score (ad esempio Reciprocal Rank Fusion, CombSUM/CombMNZ) e il re-ranking tramite cross-encoder.<sup>[\[1\]](https://systems-analysis.info/int/Hybrid_retrieval_(IT)#cite_note-1)[\[2\]](https://systems-analysis.info/int/Hybrid_retrieval_(IT)#cite_note-2)[\[3\]](https://systems-analysis.info/int/Hybrid_retrieval_(IT)#cite_note-3)</sup>

## Definizione e motivazione

*Il recupero ibrido* è una ricerca parallela o a cascata su due (o più) canali di segnali indipendenti, seguita da fusione e/o re-ranking. Le motivazioni tipiche sono: (i) superare il «divario terminologico» (sinonimi, riformulazioni), (ii) robustezza agli errori di battitura e alla morfologia, (iii) recupero di codici/identificatori specifici (dove il modello sparse è più efficace), (iv) trasferimento a nuovi domini/lingue (dove i modelli dense forniscono generalizzazione semantica).<sup>[\[4\]](https://systems-analysis.info/int/Hybrid_retrieval_(IT)#cite_note-4)[\[5\]](https://systems-analysis.info/int/Hybrid_retrieval_(IT)#cite_note-5)[\[6\]](https://systems-analysis.info/int/Hybrid_retrieval_(IT)#cite_note-6)</sup>

## Componenti della ricerca ibrida

### Lessicale (sparse)

- **Modelli classici.** TF-IDF e BM25/BM25F — metodi di base standard su indici invertiti; BM25 è fondato nel quadro probabilistico PRF e ampiamente utilizzato come primo stadio di ranking.<sup>[\[7\]](https://systems-analysis.info/int/Hybrid_retrieval_(IT)#cite_note-7)</sup>
- **Sparse apprendibili.**
  - **SPLADE / SPLADE++/v3.** Modello neurale sparse che apprende l'espansione e la pesatura dei termini tramite una testa MLM con regolarizzazione della sparsità; mostra risultati forti e buona trasferibilità (BEIR).<sup>[\[8\]](https://systems-analysis.info/int/Hybrid_retrieval_(IT)#cite_note-8)[\[9\]](https://systems-analysis.info/int/Hybrid_retrieval_(IT)#cite_note-9)[\[10\]](https://systems-analysis.info/int/Hybrid_retrieval_(IT)#cite_note-10)</sup>
  - **uniCOIL/COIL.** Liste invertite contestualizzate e la loro versione semplificata *uniCOIL*; compatibili con gli indici invertiti classici.<sup>[\[11\]](https://systems-analysis.info/int/Hybrid_retrieval_(IT)#cite_note-11)</sup>

### Semantico (dense/late‑interaction)

- **Bi‑encoder (single‑vector).** Query e documento vengono codificati da modelli vettoriali; la similarità è calcolata tramite dot‑product/MIPS. Esempi: DPR,<sup>[\[12\]](https://systems-analysis.info/int/Hybrid_retrieval_(IT)#cite_note-12)</sup> ANCE,<sup>[\[13\]](https://systems-analysis.info/int/Hybrid_retrieval_(IT)#cite_note-13)</sup> Contriever,<sup>[\[14\]](https://systems-analysis.info/int/Hybrid_retrieval_(IT)#cite_note-14)</sup> GTR,<sup>[\[15\]](https://systems-analysis.info/int/Hybrid_retrieval_(IT)#cite_note-15)</sup> E5.<sup>[\[16\]](https://systems-analysis.info/int/Hybrid_retrieval_(IT)#cite_note-16)</sup>
- **Late‑interaction (multi‑vector).** Modellano le corrispondenze a livello di token mediante interazione «tardiva»: ColBERT/ColBERTv2; il compromesso — maggiore precisione a fronte di un indice/latenza più elevati — è attenuato da ottimizzazioni ingegneristiche (PLAID, WARP).<sup>[\[17\]](https://systems-analysis.info/int/Hybrid_retrieval_(IT)#cite_note-17)[\[18\]](https://systems-analysis.info/int/Hybrid_retrieval_(IT)#cite_note-18)[\[19\]](https://systems-analysis.info/int/Hybrid_retrieval_(IT)#cite_note-19)</sup>

## Schemi di ibridizzazione e fusione dei ranking

- **Ricerca parallela e unione dei candidati.** Si ottengono liste di candidati indipendenti (sparse e dense) con i rispettivi score interni; segue la fusione dei ranking.<sup>[\[20\]](https://systems-analysis.info/int/Hybrid_retrieval_(IT)#cite_note-20)</sup>
- **RRF (Reciprocal Rank Fusion).** Tecnica robusta alle differenze di scala degli score che somma i ranghi reciproci:

${RRF}(d) = \sum\limits_{i = 1}^{m}\frac{1}{k + {rank}_{i}(d)}$, dove solitamente $k \approx 60$.<sup>[\[21\]](https://systems-analysis.info/int/Hybrid_retrieval_(IT)#cite_note-21)</sup> È supportata nei motori industriali (Elasticsearch/OpenSearch) come retriever/processore integrato.<sup>[\[22\]](https://systems-analysis.info/int/Hybrid_retrieval_(IT)#cite_note-22)[\[23\]](https://systems-analysis.info/int/Hybrid_retrieval_(IT)#cite_note-23)</sup>

- **CombSUM/CombMNZ e altri.** Funzioni classiche di «somma degli score» (con normalizzazione se necessaria).<sup>[\[24\]](https://systems-analysis.info/int/Hybrid_retrieval_(IT)#cite_note-24)[\[25\]](https://systems-analysis.info/int/Hybrid_retrieval_(IT)#cite_note-25)[\[26\]](https://systems-analysis.info/int/Hybrid_retrieval_(IT)#cite_note-26)</sup>
- **Miscela lineare pesata.**

$S(d) = \alpha \cdot S_{\text{sparse}}(d) + (1 - \alpha) \cdot S_{\text{dense}}(d)$, $\alpha \in \lbrack 0,1\rbrack$. La scelta di $\alpha$ può essere fissa o appresa (per collezione o per query).<sup>[\[27\]](https://systems-analysis.info/int/Hybrid_retrieval_(IT)#cite_note-27)</sup>

- **Normalizzazione degli score.** Per CombSUM/CombMNZ si applicano spesso min‑max, z‑score e altri metodi per allineare le scale;<sup>[\[28\]](https://systems-analysis.info/int/Hybrid_retrieval_(IT)#cite_note-28)</sup> in alternativa, RRF si basa esclusivamente sui ranghi.
- **Pesatura dinamica/adattiva.** Query routing, feature della query e modelli LTR per la selezione e la pesatura dei canali; lavori recenti mostrano che una miscela appresa semplice spesso supera RRF ed è poco sensibile alla normalizzazione.<sup>[\[29\]](https://systems-analysis.info/int/Hybrid_retrieval_(IT)#cite_note-29)</sup>

## Re-ranking e pipeline multi-stadio

I sistemi ibridi sono tipicamente costruiti come pipeline *retrieval → fusion → rerank*. Per il re-ranking si utilizzano:

- **Cross‑encoder (BERT/T5).** I più precisi ma costosi: MonoBERT/MonoT5 per il riordinamento dei top‑N candidati.<sup>[\[30\]](https://systems-analysis.info/int/Hybrid_retrieval_(IT)#cite_note-30)[\[31\]](https://systems-analysis.info/int/Hybrid_retrieval_(IT)#cite_note-31)</sup>
- **Late‑interaction come re-ranker.** La famiglia ColBERT può fungere anche da re-ranker; gli acceleratori moderni (PLAID, WARP) riducono la latenza senza perdita di qualità.<sup>[\[32\]](https://systems-analysis.info/int/Hybrid_retrieval_(IT)#cite_note-32)[\[33\]](https://systems-analysis.info/int/Hybrid_retrieval_(IT)#cite_note-33)</sup>

Il compromesso *qualità ↔ latenza/costo* è particolarmente rilevante nei contesti RAG e con SLA stringenti (si vedano le latenze di coda p95/p99).<sup>[\[34\]](https://systems-analysis.info/int/Hybrid_retrieval_(IT)#cite_note-34)</sup>

## Valutazione sui benchmark

- **BEIR.** Insieme unificato di collezioni/compiti eterogenei per la valutazione zero‑/out‑of‑domain dei retriever (ad es. TREC‑COVID, NFCorpus, NQ, HotpotQA, FiQA‑2018, DBPedia‑entity, ArguAna, Webis‑Touché‑2020, FEVER/Climate‑FEVER, Scidocs, SciFact, CQADupStack, ecc.).<sup>[\[35\]](https://systems-analysis.info/int/Hybrid_retrieval_(IT)#cite_note-35)</sup>
- **TREC Deep Learning / MS MARCO.** Risorse classiche per l'addestramento e la valutazione di retriever e re-ranker in regime di grandi dati.<sup>[\[36\]](https://systems-analysis.info/int/Hybrid_retrieval_(IT)#cite_note-36)[\[37\]](https://systems-analysis.info/int/Hybrid_retrieval_(IT)#cite_note-37)[\[38\]](https://systems-analysis.info/int/Hybrid_retrieval_(IT)#cite_note-38)</sup>
- **Metriche di qualità.** nDCG@k, Recall@k, MRR; per le prestazioni — latency p50/p95/p99, QPS; per l'esercizio — memoria/costo (CPU/GPU, indice).<sup>[\[39\]](https://systems-analysis.info/int/Hybrid_retrieval_(IT)#cite_note-39)[\[40\]](https://systems-analysis.info/int/Hybrid_retrieval_(IT)#cite_note-40)</sup>
- **Ablazioni.** Si raccomanda di isolare il contributo di ciascun canale/peso e la sensibilità ai parametri $k$ in RRF e $\alpha$ nella miscela; valutare la robustezza a riformulazioni e a shift OOD.<sup>[\[41\]](https://systems-analysis.info/int/Hybrid_retrieval_(IT)#cite_note-41)[\[42\]](https://systems-analysis.info/int/Hybrid_retrieval_(IT)#cite_note-42)</sup>

## Aspetti ingegneristici e pratiche di produzione

- **Indici e ANN.** FAISS (Flat/HNSW/IVF‑PQ), HNSW, ScaNN per MIPS/similarità coseno.<sup>[\[43\]](https://systems-analysis.info/int/Hybrid_retrieval_(IT)#cite_note-43)[\[44\]](https://systems-analysis.info/int/Hybrid_retrieval_(IT)#cite_note-44)[\[45\]](https://systems-analysis.info/int/Hybrid_retrieval_(IT)#cite_note-45)</sup>
- **Stack IR.** Lucene/Anserini/Pyserini per pipeline sparse, dense e ibride; riproducibilità immediata su BEIR.<sup>[\[46\]](https://systems-analysis.info/int/Hybrid_retrieval_(IT)#cite_note-46)[\[47\]](https://systems-analysis.info/int/Hybrid_retrieval_(IT)#cite_note-47)</sup>
- **Database vettoriali e motori di ricerca.** Qdrant, Weaviate, pgvector/PostgreSQL, Vespa, Elasticsearch/OpenSearch dispongono di modalità native di ricerca ibrida (BM25F+vector) e/o RRF/miscela lineare.<sup>[\[48\]](https://systems-analysis.info/int/Hybrid_retrieval_(IT)#cite_note-48)[\[49\]](https://systems-analysis.info/int/Hybrid_retrieval_(IT)#cite_note-49)[\[50\]](https://systems-analysis.info/int/Hybrid_retrieval_(IT)#cite_note-50)[\[51\]](https://systems-analysis.info/int/Hybrid_retrieval_(IT)#cite_note-51)[\[52\]](https://systems-analysis.info/int/Hybrid_retrieval_(IT)#cite_note-52)</sup>
- **Pattern RAG.** Architettura: **retrieval → fusion → rerank → contesto LLM** con limitazione dei token e tracciamento delle fonti.<sup>[\[53\]](https://systems-analysis.info/int/Hybrid_retrieval_(IT)#cite_note-53)</sup>
- **Aggiornamento degli indici, deduplicazione, tokenizzazione.** È importante allineare la tokenizzazione tra BM25 e il modello di vettorizzazione; calibrare gli score (normalizzazione/scalatura) prima della fusione.<sup>[\[54\]](https://systems-analysis.info/int/Hybrid_retrieval_(IT)#cite_note-54)</sup>

## Limiti e questioni aperte

- **Trasferibilità e multilingua.** I modelli dense (GTR/E5) migliorano il trasferimento, ma sono sensibili al dominio/lingua; i modelli sparse (SPLADE) sono spesso più robusti in OOD.<sup>[\[55\]](https://systems-analysis.info/int/Hybrid_retrieval_(IT)#cite_note-55)[\[56\]](https://systems-analysis.info/int/Hybrid_retrieval_(IT)#cite_note-56)</sup>
- **Integrazione con LLM e allucinazioni.** Il recupero ibrido riduce le omissioni e il rumore nei contesti RAG, ma non elimina completamente le allucinazioni; sono necessari re-ranker rigorosi e filtraggio delle fonti.<sup>[\[57\]](https://systems-analysis.info/int/Hybrid_retrieval_(IT)#cite_note-57)</sup>
- **Costo e privacy.** Archiviazione di indici multi-vector, compressione, cifratura e stack on‑prem; valutazione del TCO.
- **Tendenze.** HyDE/doc2query/PRF come espansione di documenti/query;<sup>[\[58\]](https://systems-analysis.info/int/Hybrid_retrieval_(IT)#cite_note-58)[\[59\]](https://systems-analysis.info/int/Hybrid_retrieval_(IT)#cite_note-59)</sup> apprendimento della miscela (per-query $\alpha$), late‑interaction più efficienti (PLAID/WARP), documenti lunghi e indici multi-vettore.<sup>[\[60\]](https://systems-analysis.info/int/Hybrid_retrieval_(IT)#cite_note-60)[\[61\]](https://systems-analysis.info/int/Hybrid_retrieval_(IT)#cite_note-61)</sup>

## Tabella comparativa dei metodi

Aggiornato al 2025‑09‑10 (esempio sulla collezione BEIR *trec‑covid*; nDCG@10 / Recall@100):<sup>[\[62\]](https://systems-analysis.info/int/Hybrid_retrieval_(IT)#cite_note-62)</sup>

| Metodo                           | Tipo (sparse/dense/hybrid) | Idea/modello                                 | Schema di fusione           | Re-ranker            | nDCG@10 / R@100                                                                 | Latenza (rel.) | Fonti                                                                                                                                                                                                                                                 |
|----------------------------------|----------------------------|----------------------------------------------|-----------------------------|----------------------|---------------------------------------------------------------------------------|----------------|-------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------|
| BM25                             | sparse                     | Corrispondenza esatta dei termini (PRF/BM25) | —                           | —                    | 0.595 / 0.109                                                                   | molto bassa    | <sup>[\[63\]](https://systems-analysis.info/int/Hybrid_retrieval_(IT)#cite_note-63)[\[64\]](https://systems-analysis.info/int/Hybrid_retrieval_(IT)#cite_note-64)</sup>                                                                               |
| SPLADE++ (ED)                    | sparse (learned)           | Espansione/pesatura sparsa dei termini       | —                           | —                    | 0.727 / 0.128                                                                   | bassa–media    | <sup>[\[65\]](https://systems-analysis.info/int/Hybrid_retrieval_(IT)#cite_note-65)[\[66\]](https://systems-analysis.info/int/Hybrid_retrieval_(IT)#cite_note-66)</sup>                                                                               |
| Contriever (MS MARCO FT)         | dense                      | Bi-encoder con apprendimento contrastivo     | —                           | —                    | 0.596 / 0.091                                                                   | media          | <sup>[\[67\]](https://systems-analysis.info/int/Hybrid_retrieval_(IT)#cite_note-67)[\[68\]](https://systems-analysis.info/int/Hybrid_retrieval_(IT)#cite_note-68)</sup>                                                                               |
| BGE‑base‑en‑v1.5                 | dense                      | Embedding universale di alta qualità         | —                           | —                    | 0.781 / 0.141                                                                   | media          | <sup>[\[69\]](https://systems-analysis.info/int/Hybrid_retrieval_(IT)#cite_note-69)</sup>                                                                                                                                                             |
| Cohere embed‑english‑v3.0        | dense                      | Modello di embedding testuale industriale    | —                           | —                    | 0.818 / 0.159                                                                   | media          | <sup>[\[70\]](https://systems-analysis.info/int/Hybrid_retrieval_(IT)#cite_note-70)</sup>                                                                                                                                                             |
| BM25 + dense (esempio: BM25+BGE) | hybrid                     | Recupero parallelo + fusione delle liste     | RRF (k≈60) o miscela pesata | opz.: MonoT5/ColBERT | (variabile in base all'implementazione; di norma \> del miglior canale singolo) | media          | <sup>[\[71\]](https://systems-analysis.info/int/Hybrid_retrieval_(IT)#cite_note-71)[\[72\]](https://systems-analysis.info/int/Hybrid_retrieval_(IT)#cite_note-72)[\[73\]](https://systems-analysis.info/int/Hybrid_retrieval_(IT)#cite_note-73)</sup> |

Confronto dei metodi su *trec‑covid*

Nota: l'ultima riga illustra lo schema; i valori esatti dipendono dalla scelta dell'embedding, dalla normalizzazione e dai parametri di fusione (si vedano le fonti e gli script riproducibili di Pyserini).

## Collegamenti esterni

- Pyserini / Anserini: github.com/castorini/pyserini • github.com/castorini/anserini
- FAISS: arXiv:1702.08734
- Weaviate (Hybrid search): docs.weaviate.io/weaviate/search/hybrid
- pgvector: github.com/pgvector/pgvector
- Vespa (Hybrid search tutorial): docs.vespa.ai/en/tutorials/hybrid-search.html

## Bibliografia

- Manning, C.D., Raghavan, P., Schütze, H. (2008). *Introduction to Information Retrieval*. Cambridge University Press. ISBN 978‑0521865715.
- Robertson, S., Zaragoza, H. (2009). *The Probabilistic Relevance Framework: BM25 and Beyond*. Foundations and Trends in Information Retrieval 3(4):333–389. DOI:10.1561/1500000019.
- Lin, J. et al. (2021). *Pyserini: A Python Toolkit for Reproducible IR*. SIGIR.
- Järvelин, K., Kekäläinen, J. (2002). *Cumulated Gain‑Based Evaluation of IR Techniques*. Information Retrieval 6:241–256. DOI:10.1023/A:1016043826386.
- Dean, J., Barroso, L.A. (2013). *The Tail at Scale*. CACM 56(2):74–80. DOI:10.1145/2408776.2408794.

## Note

1.  <span id="cite_note-1">[↑](https://systems-analysis.info/int/Hybrid_retrieval_(IT)#cite_ref-1) Robertson, S., Zaragoza, H. (2009). *The Probabilistic Relevance Framework: BM25 and Beyond*. Foundations and Trends in Information Retrieval, 3(4), 333–389. DOI:10.1561/1500000019.</span>
2.  <span id="cite_note-2">[↑](https://systems-analysis.info/int/Hybrid_retrieval_(IT)#cite_ref-2) Cormack, G.V., Clarke, C.L.A., Büttcher, S. (2009). *Reciprocal Rank Fusion Outperforms Condorcet and Individual Rank Learning Methods*. SIGIR 2009, 758–759. <a href="https://cormack.uwaterloo.ca/cormacksigir09-rrf.pdf" class="external text" rel="nofollow">PDF</a>.</span>
3.  <span id="cite_note-3">[↑](https://systems-analysis.info/int/Hybrid_retrieval_(IT)#cite_ref-3) Bruch, S., Gai, S., Ingber, A. (2023). *An Analysis of Fusion Functions for Hybrid Retrieval*. ACM TOIS 42(1):1–35. DOI:10.1145/3596512 • arXiv:2210.11934.</span>
4.  <span id="cite_note-4">[↑](https://systems-analysis.info/int/Hybrid_retrieval_(IT)#cite_ref-4) Manning, C.D., Raghavan, P., Schütze, H. (2008). *Introduction to Information Retrieval*. Cambridge University Press. ISBN 978‑0521865715 (см. главы о TF‑IDF, оценке и проблеме *vocabulary mismatch*).</span>
5.  <span id="cite_note-5">[↑](https://systems-analysis.info/int/Hybrid_retrieval_(IT)#cite_ref-5) Izacard, G. et al. (2022). *Unsupervised Dense Information Retrieval with Contrastive Learning (Contriever)*. TACL 10:1089–1108. arXiv:2112.09118.</span>
6.  <span id="cite_note-6">[↑](https://systems-analysis.info/int/Hybrid_retrieval_(IT)#cite_ref-6) Wang, L. et al. (2022/2024). *Text Embeddings by Weakly‑Supervised Contrastive Pre‑training (E5)*. arXiv:2212.03533.</span>
7.  <span id="cite_note-7">[↑](https://systems-analysis.info/int/Hybrid_retrieval_(IT)#cite_ref-7) Robertson, S., Zaragoza, H. (2009). *The Probabilistic Relevance Framework: BM25 and Beyond*. DOI:10.1561/1500000019.</span>
8.  <span id="cite_note-8">[↑](https://systems-analysis.info/int/Hybrid_retrieval_(IT)#cite_ref-8) Formal, T., Piwowarski, B., Clinchant, S. (2021). *SPLADE: Sparse Lexical and Expansion Model for First Stage Ranking*. arXiv:2107.05720.</span>
9.  <span id="cite_note-9">[↑](https://systems-analysis.info/int/Hybrid_retrieval_(IT)#cite_ref-9) Formal, T. et al. (2022). *Making Sparse Neural IR Models More Effective*. Findings of EMNLP. arXiv:2205.04733.</span>
10. <span id="cite_note-10">[↑](https://systems-analysis.info/int/Hybrid_retrieval_(IT)#cite_ref-10) Formal, T. et al. (2024). *SPLADE‑v3: New baselines for SPLADE*. arXiv:2403.06789.</span>
11. <span id="cite_note-11">[↑](https://systems-analysis.info/int/Hybrid_retrieval_(IT)#cite_ref-11) Lin, J., Ma, X. (2021). *A Few Brief Notes on DeepImpact, COIL, and uniCOIL*. arXiv:2106.14807.</span>
12. <span id="cite_note-12">[↑](https://systems-analysis.info/int/Hybrid_retrieval_(IT)#cite_ref-12) Karpukhin, V. et al. (2020). *Dense Passage Retrieval for Open‑Domain QA*. EMNLP. arXiv:2004.04906.</span>
13. <span id="cite_note-13">[↑](https://systems-analysis.info/int/Hybrid_retrieval_(IT)#cite_ref-13) Xiong, L. et al. (2021). *Approximate Nearest Neighbor Negative Contrastive Learning for Dense Text Retrieval (ANCE)*. ICLR. arXiv:2007.00808.</span>
14. <span id="cite_note-14">[↑](https://systems-analysis.info/int/Hybrid_retrieval_(IT)#cite_ref-14) Izacard, G. et al. (2022). TACL. arXiv:2112.09118.</span>
15. <span id="cite_note-15">[↑](https://systems-analysis.info/int/Hybrid_retrieval_(IT)#cite_ref-15) Ni, J. et al. (2021/2022). *Large Dual Encoders Are Generalizable Retrievers (GTR)*. EMNLP. arXiv:2112.07899.</span>
16. <span id="cite_note-16">[↑](https://systems-analysis.info/int/Hybrid_retrieval_(IT)#cite_ref-16) Wang, L. et al. (2022/2024). arXiv:2212.03533.</span>
17. <span id="cite_note-17">[↑](https://systems-analysis.info/int/Hybrid_retrieval_(IT)#cite_ref-17) Khattab, O., Zaharia, M. (2020). *ColBERT: Efficient and Effective Passage Search via Contextualized Late Interaction over BERT*. SIGIR. arXiv:2004.12832.</span>
18. <span id="cite_note-18">[↑](https://systems-analysis.info/int/Hybrid_retrieval_(IT)#cite_ref-18) Santhanam, K. et al. (2022). *ColBERTv2 & PLAID*. NAACL/ArXiv. arXiv:2112.01488; arXiv:2205.09707.</span>
19. <span id="cite_note-19">[↑](https://systems-analysis.info/int/Hybrid_retrieval_(IT)#cite_ref-19) Scheerer, J.L. et al. (2025). *WARP: An Efficient Engine for Multi‑Vector Retrieval*. arXiv:2501.17788.</span>
20. <span id="cite_note-20">[↑](https://systems-analysis.info/int/Hybrid_retrieval_(IT)#cite_ref-20) Lin, J. et al. (2021). *Pyserini: A Python Toolkit for Reproducible IR with Sparse and Dense Representations*. SIGIR. <a href="https://cs.uwaterloo.ca/~jimmylin/publications/Lin_etal_SIGIR2021_Pyserini.pdf" class="external text" rel="nofollow">PDF</a>.</span>
21. <span id="cite_note-21">[↑](https://systems-analysis.info/int/Hybrid_retrieval_(IT)#cite_ref-21) Cormack, G.V., Clarke, C.L.A., Büttcher, S. (2009). SIGIR. <a href="https://cormack.uwaterloo.ca/cormacksigir09-rrf.pdf" class="external text" rel="nofollow">PDF</a>.</span>
22. <span id="cite_note-22">[↑](https://systems-analysis.info/int/Hybrid_retrieval_(IT)#cite_ref-22) Elastic Docs. *Reciprocal Rank Fusion*. (доступ 2025‑09‑10). <a href="https://www.elastic.co/docs/reference/elasticsearch/rest-apis/reciprocal-rank-fusion" class="external text" rel="nofollow">elastic.co/docs/.../reciprocal-rank-fusion</a>.</span>
23. <span id="cite_note-23">[↑](https://systems-analysis.info/int/Hybrid_retrieval_(IT)#cite_ref-23) OpenSearch Docs. *Score ranker processor (RRF)*. (доступ 2025‑09‑10). <a href="https://docs.opensearch.org/3.0/search-plugins/search-pipelines/score-ranker-processor/" class="external text" rel="nofollow">docs.opensearch.org/.../score-ranker-processor/</a>.</span>
24. <span id="cite_note-24">[↑](https://systems-analysis.info/int/Hybrid_retrieval_(IT)#cite_ref-24) Fox, E.A., Shaw, J.A. (1994). *Combination of Multiple Searches*. TREC‑2, NIST SP 500‑215, 243–252. <a href="https://www.khoury.northeastern.edu/home/jaa/CSG339.06F/resources/fox94combination.pdf" class="external text" rel="nofollow">PDF</a>.</span>
25. <span id="cite_note-25">[↑](https://systems-analysis.info/int/Hybrid_retrieval_(IT)#cite_ref-25) Lee, J.H. (1997). *Analyses of Multiple Evidence Combination*. SIGIR, 267–276. DOI:10.1145/258525.258587.</span>
26. <span id="cite_note-26">[↑](https://systems-analysis.info/int/Hybrid_retrieval_(IT)#cite_ref-26) Hsu, D.F., Taksa, I. (2005). *Comparing Rank and Score Combination Methods for Data Fusion in IR*. (Tech. report). <a href="https://ccc.inaoep.mx/~villasen/bib/Hsu-FusionInIR07.pdf" class="external text" rel="nofollow">PDF</a>.</span>
27. <span id="cite_note-27">[↑](https://systems-analysis.info/int/Hybrid_retrieval_(IT)#cite_ref-27) Bruch, S., Gai, S., Ingber, A. (2023). TOIS. DOI:10.1145/3596512.</span>
28. <span id="cite_note-28">[↑](https://systems-analysis.info/int/Hybrid_retrieval_(IT)#cite_ref-28) Hsu, D.F., Taksa, I. (2005). см. выше.</span>
29. <span id="cite_note-29">[↑](https://systems-analysis.info/int/Hybrid_retrieval_(IT)#cite_ref-29) Bruch, S., Gai, S., Ingber, A. (2023). TOIS. DOI:10.1145/3596512.</span>
30. <span id="cite_note-30">[↑](https://systems-analysis.info/int/Hybrid_retrieval_(IT)#cite_ref-30) Nogueira, R., Cho, K. (2019). *Passage Re‑ranking with BERT*. arXiv:1901.04085.</span>
31. <span id="cite_note-31">[↑](https://systems-analysis.info/int/Hybrid_retrieval_(IT)#cite_ref-31) Nogueira, R., Jiang, Z., Lin, J. (2020). *Document Ranking with a Pretrained Sequence‑to‑Sequence Model (MonoT5)*. Findings of EMNLP. arXiv:2003.06713.</span>
32. <span id="cite_note-32">[↑](https://systems-analysis.info/int/Hybrid_retrieval_(IT)#cite_ref-32) Santhanam, K. et al. (2022). arXiv:2205.09707.</span>
33. <span id="cite_note-33">[↑](https://systems-analysis.info/int/Hybrid_retrieval_(IT)#cite_ref-33) Scheerer, J.L. et al. (2025). arXiv:2501.17788.</span>
34. <span id="cite_note-34">[↑](https://systems-analysis.info/int/Hybrid_retrieval_(IT)#cite_ref-34) Dean, J., Barroso, L.A. (2013). *The Tail at Scale*. CACM 56(2):74–80. DOI:10.1145/2408776.2408794.</span>
35. <span id="cite_note-35">[↑](https://systems-analysis.info/int/Hybrid_retrieval_(IT)#cite_ref-35) Thakur, N. et al. (2021). *BEIR: A Heterogeneous Benchmark for Zero‑shot Evaluation of IR Models*. NeurIPS Datasets & Benchmarks. arXiv:2104.08663.</span>
36. <span id="cite_note-36">[↑](https://systems-analysis.info/int/Hybrid_retrieval_(IT)#cite_ref-36) Craswell, N. et al. (2020). *Overview of the TREC 2019 Deep Learning Track*. arXiv:2003.07820.</span>
37. <span id="cite_note-37">[↑](https://systems-analysis.info/int/Hybrid_retrieval_(IT)#cite_ref-37) Craswell, N. et al. (2021). *Overview of the TREC 2020 Deep Learning Track*. arXiv:2102.07662.</span>
38. <span id="cite_note-38">[↑](https://systems-analysis.info/int/Hybrid_retrieval_(IT)#cite_ref-38) Bajaj, P. et al. (2016). *MS MARCO: A Human Generated MAchine Reading COmprehension Dataset*. arXiv:1611.09268.</span>
39. <span id="cite_note-39">[↑](https://systems-analysis.info/int/Hybrid_retrieval_(IT)#cite_ref-39) Järvelin, K., Kekäläinen, J. (2002). *Cumulated Gain‑Based Evaluation of IR Techniques*. Information Retrieval 6:241–256. DOI:10.1023/A:1016043826386.</span>
40. <span id="cite_note-40">[↑](https://systems-analysis.info/int/Hybrid_retrieval_(IT)#cite_ref-40) Dean, J., Barroso, L.A. (2013). CACM. DOI:10.1145/2408776.2408794.</span>
41. <span id="cite_note-41">[↑](https://systems-analysis.info/int/Hybrid_retrieval_(IT)#cite_ref-41) Bruch, S. et al. (2023). DOI:10.1145/3596512.</span>
42. <span id="cite_note-42">[↑](https://systems-analysis.info/int/Hybrid_retrieval_(IT)#cite_ref-42) Ni, J. et al. (2021/2022). arXiv:2112.07899.</span>
43. <span id="cite_note-43">[↑](https://systems-analysis.info/int/Hybrid_retrieval_(IT)#cite_ref-43) Johnson, J., Douze, M., Jégou, H. (2017). *Billion‑scale Similarity Search with GPUs (FAISS)*. arXiv:1702.08734.</span>
44. <span id="cite_note-44">[↑](https://systems-analysis.info/int/Hybrid_retrieval_(IT)#cite_ref-44) Malkov, Y., Yashunin, D. (2020). *HNSW*. IEEE TPAMI 42(4):824–836. DOI:10.1109/TPAMI.2018.2889473.</span>
45. <span id="cite_note-45">[↑](https://systems-analysis.info/int/Hybrid_retrieval_(IT)#cite_ref-45) Guo, R. et al. (2020). *ScaNN: Efficient Vector Similarity Search at Scale*. arXiv:1908.10396.</span>
46. <span id="cite_note-46">[↑](https://systems-analysis.info/int/Hybrid_retrieval_(IT)#cite_ref-46) Yang, P., Fang, H., Lin, J. (2018). *Anserini: Reproducible IR Research with Lucene*. JDIQ 10(4):1–20. DOI:10.1145/3239571.</span>
47. <span id="cite_note-47">[↑](https://systems-analysis.info/int/Hybrid_retrieval_(IT)#cite_ref-47) Lin, J. et al. (2021). SIGIR. <a href="https://cs.uwaterloo.ca/~jimmylin/publications/Lin_etal_SIGIR2021_Pyserini.pdf" class="external text" rel="nofollow">PDF</a>.</span>
48. <span id="cite_note-48">[↑](https://systems-analysis.info/int/Hybrid_retrieval_(IT)#cite_ref-48) Qdrant Docs. *Hybrid queries (RRF, DBSF)*. (доступ 2025‑09‑10). <a href="https://qdrant.tech/documentation/concepts/hybrid-queries/" class="external text" rel="nofollow">qdrant.tech/.../hybrid-queries/</a>.</span>
49. <span id="cite_note-49">[↑](https://systems-analysis.info/int/Hybrid_retrieval_(IT)#cite_ref-49) Weaviate Docs. *Hybrid search*. (доступ 2025‑09‑10). <a href="https://docs.weaviate.io/weaviate/search/hybrid" class="external text" rel="nofollow">docs.weaviate.io/weaviate/search/hybrid</a>.</span>
50. <span id="cite_note-50">[↑](https://systems-analysis.info/int/Hybrid_retrieval_(IT)#cite_ref-50) pgvector GitHub. (доступ 2025‑09‑10). <a href="https://github.com/pgvector/pgvector" class="external text" rel="nofollow">github.com/pgvector/pgvector</a>.</span>
51. <span id="cite_note-51">[↑](https://systems-analysis.info/int/Hybrid_retrieval_(IT)#cite_ref-51) Vespa Docs. *Hybrid Text Search Tutorial*. (доступ 2025‑09‑10). <a href="https://docs.vespa.ai/en/tutorials/hybrid-search.html" class="external text" rel="nofollow">docs.vespa.ai/.../hybrid-search.html</a>.</span>
52. <span id="cite_note-52">[↑](https://systems-analysis.info/int/Hybrid_retrieval_(IT)#cite_ref-52) Elastic Docs. *Reciprocal Rank Fusion*. (доступ 2025‑09‑10). <a href="https://www.elastic.co/docs/reference/elasticsearch/rest-apis/reciprocal-rank-fusion" class="external text" rel="nofollow">elastic.co/docs/.../rrf</a>.</span>
53. <span id="cite_note-53">[↑](https://systems-analysis.info/int/Hybrid_retrieval_(IT)#cite_ref-53) Lewis, P. et al. (2020). *Retrieval‑Augmented Generation for Knowledge‑Intensive NLP Tasks*. NeurIPS. arXiv:2005.11401.</span>
54. <span id="cite_note-54">[↑](https://systems-analysis.info/int/Hybrid_retrieval_(IT)#cite_ref-54) Hsu, D.F., Taksa, I. (2005). см. выше.</span>
55. <span id="cite_note-55">[↑](https://systems-analysis.info/int/Hybrid_retrieval_(IT)#cite_ref-55) Ni, J. et al. (2021/2022). arXiv:2112.07899.</span>
56. <span id="cite_note-56">[↑](https://systems-analysis.info/int/Hybrid_retrieval_(IT)#cite_ref-56) Formal, T. et al. (2021, 2022, 2024). arXiv:2107.05720; 2205.04733; 2403.06789.</span>
57. <span id="cite_note-57">[↑](https://systems-analysis.info/int/Hybrid_retrieval_(IT)#cite_ref-57) Lewis, P. et al. (2020). arXiv:2005.11401.</span>
58. <span id="cite_note-58">[↑](https://systems-analysis.info/int/Hybrid_retrieval_(IT)#cite_ref-58) Gao, L. et al. (2023). *Precise Zero‑Shot Dense Retrieval without Relevance Labels (HyDE)*. ACL. arXiv:2212.10496.</span>
59. <span id="cite_note-59">[↑](https://systems-analysis.info/int/Hybrid_retrieval_(IT)#cite_ref-59) Nogueira, R. et al. (2019). *Document Expansion by Query Prediction*. arXiv:1904.08375; *docTTTTTquery*. <a href="https://cs.uwaterloo.ca/~jimmylin/publications/Nogueira_Lin_2019_docTTTTTquery-v2.pdf" class="external text" rel="nofollow">PDF</a>.</span>
60. <span id="cite_note-60">[↑](https://systems-analysis.info/int/Hybrid_retrieval_(IT)#cite_ref-60) Santhanam, K. et al. (2022). arXiv:2205.09707.</span>
61. <span id="cite_note-61">[↑](https://systems-analysis.info/int/Hybrid_retrieval_(IT)#cite_ref-61) Scheerer, J.L. et al. (2025). arXiv:2501.17788.</span>
62. <span id="cite_note-62">[↑](https://systems-analysis.info/int/Hybrid_retrieval_(IT)#cite_ref-62) Pyserini *BEIR Regressions* (доступ 2025‑09‑10): результаты по *trec‑covid* для BM25/SPLADE/Contriever/BGE/Cohere. <a href="https://castorini.github.io/pyserini/2cr/beir.html" class="external text" rel="nofollow">castorini.github.io/pyserini/2cr/beir.html</a>.</span>
63. <span id="cite_note-63">[↑](https://systems-analysis.info/int/Hybrid_retrieval_(IT)#cite_ref-63) Robertson, S., Zaragoza, H. (2009). DOI:10.1561/1500000019.</span>
64. <span id="cite_note-64">[↑](https://systems-analysis.info/int/Hybrid_retrieval_(IT)#cite_ref-64) Pyserini BEIR. См. ссылку выше.</span>
65. <span id="cite_note-65">[↑](https://systems-analysis.info/int/Hybrid_retrieval_(IT)#cite_ref-65) Formal, T. et al. (2021, 2022). arXiv:2107.05720; 2205.04733.</span>
66. <span id="cite_note-66">[↑](https://systems-analysis.info/int/Hybrid_retrieval_(IT)#cite_ref-66) Pyserini BEIR.</span>
67. <span id="cite_note-67">[↑](https://systems-analysis.info/int/Hybrid_retrieval_(IT)#cite_ref-67) Izacard, G. et al. (2022). arXiv:2112.09118.</span>
68. <span id="cite_note-68">[↑](https://systems-analysis.info/int/Hybrid_retrieval_(IT)#cite_ref-68) Pyserini BEIR.</span>
69. <span id="cite_note-69">[↑](https://systems-analysis.info/int/Hybrid_retrieval_(IT)#cite_ref-69) Pyserini BEIR.</span>
70. <span id="cite_note-70">[↑](https://systems-analysis.info/int/Hybrid_retrieval_(IT)#cite_ref-70) Pyserini BEIR.</span>
71. <span id="cite_note-71">[↑](https://systems-analysis.info/int/Hybrid_retrieval_(IT)#cite_ref-71) Cormack et al. (2009). SIGIR. RRF.</span>
72. <span id="cite_note-72">[↑](https://systems-analysis.info/int/Hybrid_retrieval_(IT)#cite_ref-72) Bruch et al. (2023). TOIS.</span>
73. <span id="cite_note-73">[↑](https://systems-analysis.info/int/Hybrid_retrieval_(IT)#cite_ref-73) Elastic/OpenSearch RRF Docs.</span>