Source-Aware Reranking for Retrieval-Augmented Generation: A Reliability Prior Approach

# 1 Introduction
Source: [https://arxiv.org/html/2607.22584](https://arxiv.org/html/2607.22584)
Source\-Aware Reranking for Retrieval\-Augmented Generation:

A Reliability Prior Approach

Yuktha Tata Koganti, Hugo Garrido\-Lestache Belinchon

Department of Computer Science and Software Engineering

Milwaukee School of Engineering

1025 N Broadway St, Milwaukee, WI 53202

kogantiy@msoe\.edu, garrido\-lestacheh@msoe\.edu

Abstract

Standard Retrieval\-Augmented Generation pipelines rank retrieved documents by semantic similarity alone, without accounting for source provenance or credibility\. This work evaluates a simple and interpretable modification to RAG retrieval ranking that incorporates domain\-informed source reliability priors\. Each document is assigned a priorλ​\(s\)\\lambda\(s\)based on its source type, and retrieval scores are reweighted usingscore​\(q,d\)=sim​\(q,d\)×λ​\(s\)\\text\{score\}\(q,d\)=\\text\{sim\}\(q,d\)\\times\\lambda\(s\)\. The framework is evaluated against a similarity\-only baseline on a 120\-document health\-domain corpus\. In this controlled setting, source\-aware reranking improves Precision@5 from 0\.48 to 0\.72 and reduces average adversarial document retrieval under the evaluated threat model, where low\-credibility sources are identifiable via metadata\. All experiments were executed on Rosie, the high\-performance computing cluster at the Milwaukee School of Engineering, which provided the GPU\-accelerated infrastructure necessary to run the full experimental pipeline reliably and reproducibly\. These results suggest a potential mitigation strategy for source quality degradation in RAG pipelines, within the limits of the experimental setup described\.

Large language models are increasingly deployed in domains where factual accuracy carries real\-world consequences, including medicine, law, and public health\. Retrieval\-Augmented Generation addresses a fundamental limitation of static language models by grounding responses in externally retrieved documents, allowing systems to incorporate up\-to\-date and domain\-specific knowledge at inference time\. However, standard RAG pipelines retrieve documents based solely on semantic similarity between query and document embeddings\. This design treats all retrieved documents as equally valid inputs regardless of their origin, credibility, or potential for harm\.

In open\-domain or uncurated deployment settings, this can introduce a potential vulnerability\. A peer\-reviewed clinical guideline and an AI\-generated post may produce similar embedding representations if their surface content addresses the same topic, and a similarity\-only retriever has no signal to distinguish between them based on source credibility\. In adversarial settings, semantically plausible but factually misleading documents can be injected into a retrieval corpus and retrieved without penalty under a similarity\-only ranking function\.

This work addresses the problem by introducing a source reliability prior into the RAG retrieval ranking function\. The framework assigns each document a reliability prior based on its source type, multiplies retrieval similarity scores by this prior during reranking, and optionally propagates reliability signals to semantically adjacent documents through a single\-hop smoothing step\. A post\-query weight update mechanism adjusts priors based on heuristic feedback across sequential interactions\.

This work does not propose a new retrieval model\. It evaluates a simple and interpretable modification to retrieval ranking that incorporates source\-level priors, and examines whether this modification reduces retrieval of low\-credibility content in a controlled setting\. The contribution lies in systematically characterizing how source\-level priors interact with retrieval behavior across standard, adversarial, and parameter sensitivity settings\.

This paper makes three contributions\. First, a reliability\-weighted reranking formulation that operates as a post\-retrieval modification without altering the underlying language model or embedding model\. Second, a controlled adversarial injection evaluation under a specific threat model where low\-credibility documents are identifiable by source type metadata—results are interpreted within this threat model only\. Third, an empirical analysis on a 120\-document health\-domain corpus suggesting that source provenance weighting, in this controlled setting, reduces adversarial retrieval and improves precision over a similarity\-only baseline\.

## 2 Related Work

Retrieval\-Augmented Generation\.Lewis et al\. introduced RAG as a framework combining parametric language model knowledge with non\-parametric dense retrieval over document collections, demonstrating substantially reduced hallucination in open\-domain question answering\[[6](https://arxiv.org/html/2607.22584#bib.bib1)\]\. Guu et al\. extended this direction with REALM, which jointly trains the retriever and generator end\-to\-end\[[3](https://arxiv.org/html/2607.22584#bib.bib2)\]\. Both systems retrieve documents purely on the basis of embedding similarity and include no mechanism for assessing source credibility\. Our work extends this foundation by introducing a multiplicative reliability term into the retrieval ranking function\.

Dense Passage Retrieval and Indexing\.Karpukhin et al\. established dense passage retrieval using bi\-encoder architectures as an effective alternative to sparse retrieval methods\[[5](https://arxiv.org/html/2607.22584#bib.bib3)\]\. Johnson et al\. developed FAISS as a scalable approximate nearest\-neighbor index enabling efficient similarity search over large document collections\[[4](https://arxiv.org/html/2607.22584#bib.bib4)\]\. Our retrieval layer builds directly on these methods, using SentenceTransformers with FAISS for candidate retrieval before applying reliability reranking\.

Adversarial Robustness in Neural Retrieval\.Prior work has demonstrated that neural retrieval systems are susceptible to adversarial passages crafted to appear semantically relevant while containing misleading content\. Carlini et al\. and related work on corpus poisoning have shown that embedding\-based retrievers can be manipulated through injection of high\-similarity low\-quality documents\[[1](https://arxiv.org/html/2607.22584#bib.bib5)\]\. Our adversarial injection experiment directly operationalizes this threat model and measures the suppression effect of reliability priors under specific and bounded conditions\.

Trust and Provenance in Information Systems\.Research in information credibility and knowledge graph quality has established that source provenance is a meaningful signal for assessing reliability\[[2](https://arxiv.org/html/2607.22584#bib.bib6)\]\. Work on misinformation detection has demonstrated that source\-level signals provide predictive value independent of content\-level signals\[[7](https://arxiv.org/html/2607.22584#bib.bib7)\]\. Our work differs from these approaches in that it operates at the retrieval layer of an LLM pipeline rather than as a standalone classification system, and uses predefined domain\-informed priors rather than learned classifiers\.

## 3 Implementation

### 3\.1 System Architecture

The proposed framework extends the standard RAG retrieval pipeline with a source\-aware reranking layer that operates between the FAISS index and the final context window\. The system shares its embedding and retrieval components with the baseline: documents are embedded using a sentence transformer encoder, indexed via approximate nearest\-neighbor search, and retrieved as a candidate pool\. The source\-aware system then applies reliability weighting, optional neighbor smoothing, and optional post\-query weight updates before returning the final top\-kkresults\. The baseline returns top\-kkresults directly from FAISS by similarity rank\. Both systems are evaluated using identical metrics on identical query sets\.

The core ranking modification—multiplying similarity by a source prior—is intentionally simple\. The contribution of this work is not algorithmic novelty but an empirical evaluation of whether this simple modification produces meaningful retrieval differences in a controlled setting, and an honest characterization of its limitations and boundary conditions\.

![Refer to caption](https://arxiv.org/html/2607.22584v1/images/Architecture.png)Figure 1:System architecture overview\. The left branch shows the baseline similarity\-only pipeline\. The right branch shows the source\-aware reranking pipeline with reliability weighting, optional neighbor smoothing, and optional post\-query weight update\. Both branches share embedding and retrieval components and are evaluated on identical query sets\.
### 3\.2 Corpus and Source Taxonomy

The evaluation corpus consists of 120 documents in the health and medical information domain, spanning seven source types\. The health domain was selected because source reliability carries clear real\-world meaning in this context—the distinction between a peer\-reviewed clinical study and an AI\-generated health post carries practical significance for downstream answer quality\. The corpus includes 10 adversarially crafted documents with source typesai\_generatedoruser\_generated, each written to be semantically close to the evaluation query topics while containing misleading or false medical claims\. These documents are known only to the evaluation layer and are not flagged during retrieval\.

### 3\.3 Baseline Retrieval

The baseline system embeds all documents using the all\-MiniLM\-L6\-v2 sentence transformer\[[8](https://arxiv.org/html/2607.22584#bib.bib8)\]and indexes them using FAISS with inner product similarity over L2\-normalized vectors, equivalent to cosine similarity\. For a queryqqand documentdd, the baseline retrieval score is:

scorebaseline​\(q,d\)=sim​\(q,d\)\\text\{score\}\_\{\\text\{baseline\}\}\(q,d\)=\\text\{sim\}\(q,d\)\(1\)
The top\-kkdocuments by this score form the baseline retrieved set with no further modification\.

### 3\.4 Reliability\-Weighted Reranking

The source\-aware system retrieves a candidate pool of3​k3kdocuments by similarity, then reranks using a reliability\-weighted score:

scoreSA​\(q,d\)=sim​\(q,d\)×λ​\(sd\)\\text\{score\}\_\{\\text\{SA\}\}\(q,d\)=\\text\{sim\}\(q,d\)\\times\\lambda\(s\_\{d\}\)\(2\)
wheresds\_\{d\}is the source type of documentddandλ​\(sd\)\\lambda\(s\_\{d\}\)is a predefined domain\-informed prior\. The expanded candidate pool of3​k3kis necessary to allow reliability weighting to meaningfully reorder results—reranking only the final top\-kkwould produce no change in ranking order\. Initial priors are shown in Table[1](https://arxiv.org/html/2607.22584#S3.T1)\.

Table 1:Source type reliability priorsλ​\(s\)\\lambda\(s\)\.These values represent domain\-informed assumptions about relative source credibility in the health domain\. They are not learned from data and are not claimed to be optimal\. The sensitivity of results to these values is examined empirically in Section[5\.4](https://arxiv.org/html/2607.22584#S5.SS4)\.

### 3\.5 Optional Neighbor Smoothing

After reliability weighting, an optional single\-hop smoothing step adjusts each candidate document’s effective reliability based on its nearest neighbors in the index:

λsmoothed​\(d\)=λ​\(d\)\+α​∑n∈𝒩​\(d\)sim​\(d,n\)⋅λ​\(n\)\\lambda\_\{\\text\{smoothed\}\}\(d\)=\\lambda\(d\)\+\\alpha\\sum\_\{n\\in\\mathcal\{N\}\(d\)\}\\text\{sim\}\(d,n\)\\cdot\\lambda\(n\)\(3\)
where𝒩​\(d\)\\mathcal\{N\}\(d\)denotes the top\-3 nearest neighbors ofdd,α=0\.10\\alpha=0\.10is a fixed smoothing coefficient, and the result is clipped to\[0\.10,1\.00\]\[0\.10,\\,1\.00\]\. This step is a heuristic adjustment, not a principled propagation algorithm\. On the current corpus its effect is marginal; it is included as a structural extension whose benefit may increase with corpus size\.

### 3\.6 Optional Post\-Query Weight Update

After each query, the system optionally updates source type weights using a heuristic feedback signal:

λt\+1​\(s\)=clip​\(λt​\(s\)\+η​δt,0\.10,1\.00\)\\lambda\_\{t\+1\}\(s\)=\\text\{clip\}\\\!\\left\(\\lambda\_\{t\}\(s\)\+\\eta\\,\\delta\_\{t\},\\;0\.10,\\;1\.00\\right\)\(4\)
whereη=0\.05\\eta=0\.05is the update rate andδt\\delta\_\{t\}is a simulated feedback signal based on ground\-truth relevance labels\. Specifically:δ=\+0\.3\\delta=\{\+\}0\.3if the source type matches expected relevant types,δ=−0\.5\\delta=\{\-\}0\.5if the document is adversarial, andδ=−0\.1\\delta=\{\-\}0\.1otherwise\. This feedback is oracle\-based—it uses ground\-truth labels rather than real user signals—and results from this component should be interpreted as demonstrating the behavior of the update rule under ideal conditions, not as evidence of deployable adaptive learning\.

### 3\.7 Computational Infrastructure

All experiments in this work were executed and validated on Rosie, the high\-performance computing cluster operated by the Milwaukee School of Engineering\. The repeated embedding and indexing workload across baseline, source\-aware, adversarial, sensitivity, and adaptive experimental runs would have been prohibitively slow on standard laptop hardware\. Rosie’s GPU\-accelerated nodes allowed the full pipeline to complete in a single session with consistent, reproducible results\. Its stable, dependency\-isolated Linux environment ensured identical software conditions across all runs, and version control via Git enabled every reported result to be traced to a specific experimental configuration\. The breadth of evaluation in this paper was only practical because of the infrastructure Rosie provided\.

## 4 Experimentation

### 4\.1 Evaluation Setup

Ten evaluation queries were constructed across five medical topics: vaccine safety, drug interactions, cancer screening, antibiotic resistance, and treatment efficacy\. Each query is paired with a set of relevant document IDs based on topical match, defined independently of source type\. Retrieval is evaluated atk=5k=5for all systems using the following metrics: Precision@5 \(proportion of top\-5 results that are relevant\), NDCG@5 \(normalized discounted cumulative gain using relevance labels as grades\), MRR \(reciprocal rank of the first relevant document\), and adversarial hit rate \(proportion of top\-5 results that are adversarial documents\)\. Average reliability score and high\-reliability ratio are reported separately as secondary descriptive metrics directly tied to the method’s design\.

### 4\.2 Adversarial Injection Protocol

All 10 adversarial documents were injected into the FAISS index alongside the 110 clean documents for the adversarial experiment\. Both systems were queried over all 10 evaluation queries with the full poisoned index\. The adversarial documents were designed to be semantically close to query topics, representing a hard case for similarity\-based retrieval\. The adversarial hit rate measures how frequently these documents appear in the final top\-5 results\.

### 4\.3 Adaptive Learning Protocol

The adaptive experiment processes all 10 queries sequentially, updating weights after each query using the heuristic signal described in Section 3\.6\. Weight history is logged after every query\. Convergence is assessed as mean absolute weight change over the final three queries falling below 0\.01\.

## 5 Results and Analysis

### 5\.1 Retrieval Quality

Table[2](https://arxiv.org/html/2607.22584#S5.T2)presents average retrieval metrics across all 10 queries for both systems\.

Table 2:Retrieval metrics averaged across 10 queries, baseline vs\. source\-aware\.In this controlled setting, source\-aware reranking improves Precision@5 from 0\.48 to 0\.72, a 50% relative improvement, indicating that reliability weighting consistently surfaces topically relevant documents\. The source\-aware system achieves MRR of 1\.000 across all queries, meaning the first retrieved document is always relevant across this query set\. The consistently perfect MRR suggests that the evaluation queries are relatively well\-aligned with relevant documents in the corpus, which may limit the ability to differentiate systems on this metric\. Precision@5 across individual queries ranged from 0\.40 to 1\.00 for the source\-aware system and from 0\.40 to 0\.80 for the baseline, indicating consistent directional improvement across queries though the magnitude varies\. No formal significance testing was performed given the query count of 10, and these results should be treated as exploratory rather than statistically confirmed\.

Secondary reliability metrics are reported for descriptive purposes: the source\-aware system achieved an average reliability score of 0\.863 versus 0\.737 for the baseline\. These metrics are directly tied to the method’s design and do not constitute independent evidence of retrieval quality improvement\. Figure[4](https://arxiv.org/html/2607.22584#S5.F4)further shows that adding neighbor diffusion on top of source\-aware reranking consistently matches or exceeds the source\-aware system on reliability and adversarial suppression across most queries\.

![Refer to caption](https://arxiv.org/html/2607.22584v1/images/bar_reliability_comparison.png)Figure 2:Average source reliability score per query, baseline vs\. source\-aware\. Reported as a secondary descriptive metric\.![Refer to caption](https://arxiv.org/html/2607.22584v1/images/ndcg_comparison.png)Figure 3:NDCG@5 per query, baseline vs\. source\-aware retrieval\.![Refer to caption](https://arxiv.org/html/2607.22584v1/images/diffusion_comparison.png)Figure 4:Comparison of Baseline, Source\-Aware, and Source\-Aware with neighbor diffusion across Precision@5, adversarial hit rate, average reliability, and NDCG@5 per query\. SA\+Diffusion consistently matches or exceeds Source\-Aware on reliability and adversarial suppression across most queries\.
### 5\.2 Adversarial Robustness

Table[3](https://arxiv.org/html/2607.22584#S5.T3)presents adversarial hit rates under the injection protocol\. The baseline retrieves at least one adversarial document in 9 of 10 queries, with an average adversarial hit rate of 0\.32\. Source\-aware reranking reduces this to an average of 0\.28, a 12\.5% relative reduction\. In 6 of 10 queries, source\-aware retrieval either eliminates or ties adversarial exposure\. However, two queries \(q006 and q010\) showed a slight increase in adversarial hit rate under the source\-aware system, indicating that reliability weighting is not uniformly beneficial and can in some cases surface adversarial documents that rank higher after reweighting displaces other low\-reliability clean documents\.

Table 3:Adversarial document hit rate per query topic\.This outcome is partially consistent with the experimental design: adversarial documents carry source types with low reliability priors, and the multiplicative penalty suppresses them in most cases when competing against higher\-reliability documents on the same topic\. The exceptions at q006 and q010 suggest that topic\-specific corpus structure can create conditions where the reranking interacts with document distribution in ways that are not fully controlled by the prior alone\. This should not be interpreted as robustness against adversarial content in general—an adversary aware of the prior system could assign high\-reliability source labels to misleading content, bypassing the penalty entirely\. The contribution of this experiment is to operationalize and measure the suppression effect under this specific and limited threat model\.

![Refer to caption](https://arxiv.org/html/2607.22584v1/images/adversarial_hit_rate.png)Figure 5:Adversarial document hit rate by query topic, baseline vs\. source\-aware\. Under the evaluated threat model, source\-aware reranking reduces average adversarial retrieval from 0\.32 to 0\.28, with mixed results across individual topics\.![Refer to caption](https://arxiv.org/html/2607.22584v1/images/spoofed_vs_naive_adversarial.png)Figure 6:Spoofed vs\. naive adversarial document retrieval rates across query topics\. Spoofed documents, which carry artificially elevated source labels, are harder to suppress and expose the boundary conditions of the reliability prior approach\.
### 5\.3 Adaptive Weight Evolution

Table[4](https://arxiv.org/html/2607.22584#S5.T4)shows initial and final reliability weights after 10 sequential queries under oracle feedback simulation\.

Table 4:Adaptive weight evolution under oracle feedback simulation\.Government sources receive the largest positive update \(\+\+0\.095\), followed by peer\-reviewed sources \(\+\+0\.050\), organization reports \(\+\+0\.030\), and established news \(\+\+0\.025\), consistent with their appearance in relevant top\-5 results across queries\. Source types never retrieved in the top\-5 receive no update—the system correctly makes no inference about sources it has not observed\. The mean absolute weight change over the final three queries was below the 0\.01 convergence threshold, indicating stable behavior within 10 interactions\. These results are produced under simulated oracle feedback and reflect the theoretical behavior of the update rule rather than evidence of deployed adaptive performance\.

![Refer to caption](https://arxiv.org/html/2607.22584v1/images/weight_evolution.png)Figure 7:Reliability weight evolution across 10 sequential queries under oracle feedback simulation\. Government and peer\-reviewed sources receive the largest positive updates\. Source types not retrieved in top\-5 receive no update\.![Refer to caption](https://arxiv.org/html/2607.22584v1/images/weight_delta_bar.png)Figure 8:Change in reliability weight \(final minus initial\) per source type after 10 queries\.
### 5\.4 Sensitivity Analysis

Table[5](https://arxiv.org/html/2607.22584#S5.T5)reports retrieval metrics across three weight configurations evaluated on a held\-out set of four queries \(q007–q010\)\.

Table 5:Sensitivity analysis across weight configurations on test queries \(q007–q010\)\. Baseline values are identical across configurations as the baseline does not use reliability weights\.The compressed configuration, which uses a narrower spread between high and low reliability priors \(0\.25\), produces only a modest improvement in Precision@5 \(\+\+0\.05\)\. The current and aggressive configurations, with spreads of 0\.65 and 0\.89 respectively, both achieve Precision@5 of 0\.80 \(\+\+0\.30 over baseline\)\. MRR is 1\.000 across all three configurations, indicating that the first retrieved result is always relevant regardless of weight spread\. The marginal difference between current and aggressive configurations suggests diminishing returns beyond a spread of 0\.65, and that the current configuration represents a practical operating point\. These results confirm that prior spread is the primary driver of retrieval improvement and that the current configuration is not a fragile choice\.

![Refer to caption](https://arxiv.org/html/2607.22584v1/images/sensitivity_analysis.png)Figure 9:Precision@5 and NDCG@5 across compressed, current, and aggressive weight configurations on test queries \(q007–q010\)\. The current and aggressive configurations both outperform the compressed baseline by a substantial margin\.

## 6 Challenges

The central limitation of this work is that reliability priors are manually defined rather than learned from data\. The values in Table[1](https://arxiv.org/html/2607.22584#S3.T1)represent reasonable domain assumptions but have not been empirically validated against an independent credibility ground truth\.

The adversarial evaluation uses explicitly labeled adversarial documents assigned to low\-reliability source types, meaning the method’s partial success in suppressing these documents is structurally expected given the experimental design\. A more rigorous adversarial evaluation would use documents whose source type is unknown or misrepresented, which would constitute a harder and more realistic threat model\. The cases at q006 and q010 where source\-aware performed worse than baseline highlight the limits of a purely prior\-based defense and point toward the need for content\-level signals in future work\.

The adaptive weight update relies on oracle feedback derived from ground\-truth relevance labels\. This is not a deployable feedback mechanism\. Real deployment would require implicit feedback signals such as user engagement, downstream answer quality, or expert annotation, none of which were available in this study\.

The corpus of 120 documents is domain\-specific and small relative to production RAG corpora\. With only 10 evaluation queries, no formal statistical significance testing was performed\. Results from this scale should be treated as exploratory and should not be extrapolated to larger or multi\-domain settings without further evaluation\.

## 7 Future Work

Several directions follow from this work\. Reliability priors could be estimated from large\-scale credibility datasets or learned jointly with the retrieval model, removing the dependency on manual configuration\. The adaptive update mechanism could be extended to incorporate real implicit feedback signals such as user dwell time or downstream answer quality from an evaluator model\. The single\-hop neighbor smoothing could be extended to multi\-hop graph propagation, which may provide stronger reliability signal in larger corpora where neighborhood structure is richer\.

A more robust adversarial evaluation should test against a harder threat model where adversaries can spoof source metadata, assigning high\-reliability labels to misleading content\. The per\-query variability observed in this study, particularly the cases where source\-aware retrieval performed worse than baseline, suggests that a content\-level credibility signal is needed to complement the source\-level prior\.

Evaluation on established benchmarks such as BEIR would allow direct comparison against prior retrieval systems across multiple domains and would establish whether the observed improvements generalize beyond this controlled setting\. Scaling the corpus and query set would also enable formal statistical significance testing, which was not possible at the current evaluation scale\.

## 8 Conclusion

This work introduced a source\-aware reranking framework for retrieval\-augmented generation that assigns domain\-informed reliability priors to documents and reweights retrieval scores by source type\. In a controlled evaluation on a 120\-document health\-domain corpus executed on Rosie, MSOE’s high\-performance computing cluster, the framework improves Precision@5 from 0\.48 to 0\.72 and reduces average adversarial document retrieval from 0\.32 to 0\.28 relative to a similarity\-only baseline\. MRR improves to 1\.000 across all queries, meaning the highest\-ranked retrieved document is always relevant\. Sensitivity analysis confirms that the current weight configuration is not a fragile choice, with both the current and aggressive spread configurations producing consistent improvements over the compressed baseline\. These results suggest that lightweight source provenance weighting is a viable and interpretable post\-retrieval modification that does not require changes to the underlying language model or embedding model\. The method’s limitations—manual priors, oracle feedback, small\-scale evaluation, and a specific threat model—are acknowledged explicitly and point to concrete directions for future work\.

## References

- \[1\]N\. Carlini, M\. Jagielski, C\. A\. Choquette\-Choo, D\. Paleka, W\. Pearce, H\. Anderson, A\. Terzis, K\. Thomas, and F\. Tramèr\(2024\)Poisoning web\-scale training datasets is practical\.InIEEE Symposium on Security and Privacy,Cited by:[§2](https://arxiv.org/html/2607.22584#S2.p3.1)\.
- \[2\]\(2014\)Knowledge vault: a web\-scale approach to probabilistic knowledge fusion\.InACM SIGKDD International Conference on Knowledge Discovery and Data Mining,Cited by:[§2](https://arxiv.org/html/2607.22584#S2.p4.1)\.
- \[3\]K\. Guu, K\. Lee, Z\. Tung, P\. Pasupat, and M\. Chang\(2020\)REALM: retrieval\-augmented language model pre\-training\.InInternational Conference on Machine Learning,Cited by:[§2](https://arxiv.org/html/2607.22584#S2.p1.1)\.
- \[4\]J\. Johnson, M\. Douze, and H\. Jégou\(2019\)Billion\-scale similarity search with GPUs\.IEEE Transactions on Big Data\.Cited by:[§2](https://arxiv.org/html/2607.22584#S2.p2.1)\.
- \[5\]V\. Karpukhin, B\. Oguz, S\. Min, P\. Lewis, L\. Wu, S\. Edunov, D\. Chen, and W\. Yih\(2020\)Dense passage retrieval for open\-domain question answering\.InProceedings of the 2020 Conference on Empirical Methods in Natural Language Processing,Cited by:[§2](https://arxiv.org/html/2607.22584#S2.p2.1)\.
- \[6\]P\. Lewis, E\. Perez, A\. Piktus, F\. Petroni, V\. Karpukhin, N\. Goyal, H\. Küttler, M\. Lewis, W\. Yih, T\. Rocktäschel, S\. Riedel, and D\. Kiela\(2020\)Retrieval\-augmented generation for knowledge\-intensive NLP tasks\.InAdvances in Neural Information Processing Systems,Cited by:[§2](https://arxiv.org/html/2607.22584#S2.p1.1)\.
- \[7\]V\. Pérez\-Rosas, B\. Kleinberg, A\. Lefevre, and R\. Mihalcea\(2018\)Automatic detection of fake news\.InProceedings of the 27th International Conference on Computational Linguistics,Cited by:[§2](https://arxiv.org/html/2607.22584#S2.p4.1)\.
- \[8\]N\. Reimers and I\. Gurevych\(2019\)Sentence\-BERT: sentence embeddings using siamese BERT\-networks\.InProceedings of the 2019 Conference on Empirical Methods in Natural Language Processing,Cited by:[§3\.3](https://arxiv.org/html/2607.22584#S3.SS3.p1.2)\.