Aethel: A Reproducible Graph-Retrieval Framework for Multi-Hop Financial Diligence

Krish Sapru

REVIEW 3 major objections 3 minor 12 references

This paper measures where graph-based retrieval helps and where it does not: in small candidate pools it improves multi-hop hit rate, but on a 4,123-chunk financial corpus it fails to beat BM25, and a fusion of the two yields no statistical

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

T0 review · deepseek-v4-flash

2026-08-01 15:06 UTC pith:MZVT2NNC

load-bearing objection Sound open-corpus retrieval experiments, but the paper's central claim that graph-RAG's advantage over dense shrinks at scale is backwards — the data show the opposite. the 3 major comments →

arxiv 2607.24826 v1 pith:MZVT2NNC submitted 2026-07-20 cs.IR cs.MA

Aethel: A Reproducible Graph-Retrieval Framework for Multi-Hop Financial Diligence

Krish Sapru This is my paper

classification cs.IR cs.MA

keywords graph retrievalmulti-hop question answeringpersonalized PageRankfinancial diligenceBM25 baselinedense retrievalentity coreferencereproducible evaluation

verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The paper argues that graph-based retrieval, which propagates relevance through entity–passage links via Personalized PageRank, improves multi-hop question answering in small candidate pools, but that this advantage attenuates at open-corpus scale on keyword-dense financial text. On a 4,123-chunk corpus of real financial disclosures, the graph retriever beats a dense bi-encoder on multi-hop recall (HR@5 0.600 vs 0.450) but not a strong BM25 baseline (0.700). Fusing BM25 with the graph retriever yields only a small MRR gain that is statistically indistinguishable from zero (95% CI [-0.163, +0.192]). The paper's central contribution is this honest, reproducible measurement of where graph retrieval helps and where it does not, together with a scale-sensitivity analysis showing dense retrieval collapses as the corpus grows.

Core claim

The central empirical finding is that the multi-hop retrieval advantage of graph retrieval over dense retrieval, clearly visible in small candidate pools, is substantially attenuated at open-corpus scale on keyword-dense financial text, and graph retrieval does not surpass a strong BM25 baseline. Specifically, on a 4,123-chunk corpus of real financial disclosures with 40 manually annotated queries, the coreference-aware graph retriever (Aethel-NER3) achieves multi-hop HR@5 of 0.600 versus BM25's 0.700 and Dense's 0.450. Reciprocal Rank Fusion of BM25 and Aethel-NER3 matches BM25's HR@1 and slightly edges its MRR (0.479 vs 0.468), but a paired bootstrap shows the difference is not statistical

What carries the argument

The framework models a corpus as a bipartite graph with entity nodes and passage nodes, then runs Personalized PageRank (PPR), seeded by query entities, to rank passages by steady-state probability. The Bipartite Coreference Teleportation (BCT) layer expands the seed vector with entity aliases and substring-overlap matches, improving hit-rate coverage at the cost of top-1 precision. This machinery is what lets relevance diffuse through shared entities and intermediate bridges, providing interpretable multi-hop evidence paths.

Load-bearing premise

The open-corpus financial benchmark rests on only 40 queries annotated by a single annotator, with no inter-annotator agreement reported, and the scale-sensitivity finding depends on the 20 multi-hop queries and the decision to always retain gold passages in subsamples.

What would settle it

A larger, multi-annotator financial benchmark (e.g., 200+ queries) comparing BM25, Aethel-NER3, and their fusion would settle the central claim: if the graph retriever's multi-hop HR@5 meets or exceeds BM25, or if the fusion MRR gain becomes statistically significant, the paper's conclusion that graph retrieval does not beat a strong lexical baseline would be directly contradicted.

Watch this falsifier — get emailed when new claim-graph text bears on it.

If this is right

If the central finding holds, graph-based retrieval should not be assumed to beat lexical baselines in open-corpus financial settings; BM25 remains a strong default for keyword-dense queries.
The scale-sensitivity result implies that dense retrieval's performance in small benchmarks may overstate its production utility, as it degrades sharply with corpus size.
The graph retriever's graceful degradation relative to dense retrieval suggests a role for graph methods as a complement in large corpora, even if they do not lead in absolute recall.
The entity-index quality dependence indicates that graph retrieval's value is gated by the reliability of entity extraction and linking, not just the retrieval algorithm.
The statistically null fusion result suggests that combining lexical and graph signals may require more sophisticated fusion or larger test sets to detect any real gain.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

A natural extension is to test whether adaptive damping or seed restriction in PPR, as the paper suggests, could prevent mass over-diffusion and recover top-of-ranking signal at larger scales.
The paper's finding that the graph retriever beats dense retrieval but not BM25 on financial text may generalize to other keyword-dense, proper-noun-heavy domains, such as legal or medical records, where exact lexical matches are strong cues.
If a larger multi-annotator benchmark were built, the ordering between BM25 and graph retrieval might shift for queries that require genuine relational reasoning rather than keyword anchoring; the current 20 multi-hop queries may not capture the hardest cases.
The single-annotator limitation suggests that the reported effect sizes should be treated as provisional; a sensitivity analysis with multiple annotators would clarify whether the BM25 dominance is robust.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit.

Desk Editor's Note

Sound open-corpus retrieval experiments, but the paper's central claim that graph-RAG's advantage over dense shrinks at scale is backwards — the data show the opposite.

read the letter

Colleague — this one is a useful read with a real problem. The authors release a reproducible open-corpus financial retrieval benchmark, report their limitations honestly, and the scaling experiment (Figure 3) is genuinely interesting. But the headline conclusion is contradicted by their own numbers. In the closed-pool Table 1, graph retrieval does not beat dense on MuSiQue (Dense HR@5 96.0 vs PPR 87.5, BCT 88.5) and ties on 2Wiki (99.5 vs 99.5, BCT 100). So there is no 'clear multi-hop advantage' in small pools. At open-corpus scale, Aethel-NER3 beats Dense 0.600 vs 0.450 — a bigger gap than any closed-pool edge. Figure 3 shows the same: dense leads at 100 chunks, graph passes it at 500, and the gap widens to full scale. So the graph advantage does not attenuate at scale; it only appears at scale. What does attenuate is any hope of beating BM25, which was never in the closed-pool comparison. The abstract and Section 6.3 conflate 'graph doesn't beat BM25' with 'graph's advantage over dense shrinks.' That's a serious misreading, not just a wording slip.

The good parts: the 4,123-chunk corpus is real SEC/earnings material, labels were frozen before retrieval, they report bootstrap CIs on the fusion difference, disclose the single annotator, and release code. The finding that dense retrieval collapses by 0.39 HR@5 as distractors grow while BM25 only drops 0.10 is a solid, useful observation for anyone building RAG systems. The BCT coreference layer is modest but described cleanly.

Beyond the central claim, the open-corpus benchmark is small (N=40 queries, N=20 for multi-hop), so the ordering is not tight; the fusion CI is wide. The specialist swarm is not evaluated, as the paper admits, so call it architecture. The graph index quality depends heavily on NER filtering, and the failed variant is disclosed, which is good.

I would send this to peer review, but the authors need to fix the central claim before it's published. The experiments support a different and equally interesting story: graph retrieval scales better than dense on this kind of keyword-dense financial text, but a well-tuned BM25 is still the baseline to beat. If they correct the narrative, the paper is a useful contribution to applied RAG evaluation. I'd bring it to reading group as a case study in how easy it is to tell yourself the wrong story about your own scaling experiment.

Referee Report

3 major / 3 minor

Summary. The paper presents Aethel, a graph-retrieval framework that couples bipartite Personalized PageRank (PPR) retrieval with a coreference-aware BCT layer and an agent swarm, targeting multi-hop financial diligence. It evaluates the retrieval layer on 200-question samples from MuSiQue and 2WikiMultiHopQA under closed pools, and on a 4,123-chunk financial corpus with 40 manually annotated queries (20 multi-hop). The reported central finding is that graph-RAG's multi-hop advantage, visible in small pools, is attenuated at open-corpus scale, and that the graph retriever does not beat a strong BM25 baseline. The paper also reports scale-sensitivity results and a fusion analysis with a non-significant MRR gain.

Significance. The work is valuable for its reproducibility and candor: the authors release evaluation code, report bootstrap confidence intervals, disclose single-annotator labels, stress the closed-pool nature of the small-pool results, and openly report a failed NER graph variant. These are exemplary practices for an empirical systems paper. However, the central empirical claim as stated in the abstract and Section 6.3 is not supported by the paper's own data: the graph retriever does not show a clear multi-hop advantage over the dense baseline in the closed pools, while it does show a larger advantage at open-corpus scale. If the claim is reframed—graph-over-dense advantage emerges at scale, while a lexical baseline remains strongest—the paper makes a useful, honest contribution to retrieval evaluation. But as written, the interpretation needs correction.

major comments (3)

[Abstract, §6.3, §7] The central scale-attenuation claim is contradicted by Tables 1 and 2 and Figure 3. In closed pools (Table 1), Dense beats the graph retrievers on MuSiQue (HR@5: Dense 96.0, Graph PPR 87.5, BCT 88.5) and ties on 2WikiMultiHopQA (99.5 vs. 99.5, with BCT at 100.0). In the open-corpus financial evaluation (Table 2), Aethel-NER3 beats Dense on multi-hop HR@5 by 0.150 (0.600 vs. 0.450), a larger margin than any closed-pool graph-vs-dense edge. Figure 3 confirms the graph advantage over dense is tied at 100 chunks and emerges only at 500 chunks, widening to 0.150 at full scale. Thus the graph-over-dense advantage is not attenuated at scale; it emerges at scale. If the 'advantage' refers instead to a lexical baseline, there is no closed-pool BM25 comparison, so 'attenuated' is unsupported. Please rewrite the central finding to distinguish the graph-over-dense trend from the graph-vs-BM25 orderi
[§5.2, Limitations, Table 2] The open-corpus benchmark rests on 40 queries (20 multi-hop) labeled by a single annotator, with no inter-annotator agreement reported. This is disclosed, but it is load-bearing because the comparative ordering BM25 > Aethel-NER3 > Dense (multi-hop HR@5 0.700/0.600/0.450) is the empirical basis for the paper's central conclusion. At N=20, no confidence intervals or significance tests are given for the HR@5 differences; the only CI reported (fusion MRR +0.011, [−0.163,+0.192]) crosses zero. The ordering could easily change with more queries or multiple annotators. The abstract and Section 6.3 should either present these results explicitly as preliminary/hypothesis-generating, or the authors should add a larger multi-annotator benchmark before claiming a central empirical finding.
[§6.4, Figure 3] In the scale-sensitivity experiment, gold passages are always retained in every subsample. This makes the curves a measure of distractor tolerance conditional on gold being present, not of degradation in true open retrieval, where the retriever must also locate the gold passages among all chunks. Further, no uncertainty or significance testing is reported for the NER3-vs-Dense gap: at 500 chunks the difference is 0.010 (0.670 vs. 0.660), and at full scale the difference is 0.150 with only 20 queries. The claim that the margin 'widens monotonically' is not established by the plotted bands. Please report paired tests or bootstrap CIs for the HR@5 differences, and clarify how gold retention affects interpretation.

minor comments (3)

[§5.3] The script name 'evaluate aethel.py' contains a space; this appears to be a typo for 'evaluate_aethel.py'. Please correct it in the text and in the released artifacts.
[Abstract / §3] The abstract foregrounds the BCT coreference layer, but the open-corpus evaluation uses Aethel-NER3, which does not include BCT; BCT integration is left as future work (Limitations, fourth point). Please clarify in the abstract and contribution list that BCT is validated only in the closed-pool experiments, so readers do not infer that the financial results exercise the paper's novel component.
[§6.3 / Table 2] Table 2 reports R@5 values, but the text focuses almost entirely on HR@5 and MRR. A brief sentence explaining the information added by R@5 (or removing it if redundant with HR@5) would improve readability.

Circularity Check

0 steps flagged

No circularity: the retrieval components are tested against external benchmarks with frozen labels and untuned parameters; the paper's main weakness is internal inconsistency, not circular reasoning.

full rationale

I find no circular step in Aethel's derivation chain. The graph-retrieval algorithm (bipartite PPR with alpha=0.15, iteration cap I=20) is a standard, parameter-free construction compared against external benchmarks (MuSiQue, 2WikiMultiHopQA) and against BM25 and a dense bi-encoder on a separately annotated 4,123-chunk financial corpus. The open-corpus gold labels were frozen before any retriever ran, all systems searched the same index, and no query template or hyperparameter was tuned on the evaluation sets; RRF k=60 is the canonical value cited from Cormack et al. (2009). The BCT layer is not defined in terms of the HR@k metric: alias expansion could in principle reduce or leave unchanged hit rate, and the paper reports it as an empirical tradeoff (coverage gain at precision@1 cost), so the result is not true by construction. The paper also discloses its limitations: BCT is only in the standalone evaluator, the open-corpus benchmark uses 40 single-annotator queries, and the fusion MRR difference is explicitly reported as not statistically distinguishable from zero. The only serious concern is that the abstract's central claim ('graph-RAG's multi-hop advantage ... substantially attenuated at open-corpus scale') is hard to reconcile with the paper's own tables and Figure 3: in the closed pool the graph retriever does not clearly beat dense retrieval, while at open-corpus scale Aethel-NER3 beats Dense by 0.150 HR@5. That is an internal-consistency or data-interpretation problem, not circularity, and does not raise the circularity score.

Axiom & Free-Parameter Ledger

4 free parameters · 4 axioms · 2 invented entities

The quantitative claims rest on a small number of standard free parameters (α=0.15, I=20, k=60) that are canonical and not fitted, so the fitted-parameter burden is low. The larger burdens are the single-annotator 40-query open corpus, the hand-crafted entity-index cleaning procedure, and the unvalidated agent swarm. The paper discloses all of these.

free parameters (4)

PPR damping factor α = 0.15
Standard canonical PageRank value, stated in Section 4.3 and not tuned per dataset; not fitted to the reported results. Low risk.
PPR power-iteration cap I = 20
Chosen in Section 4.3 as sufficient for convergence; not fitted to the data.
RRF fusion constant k = 60
Canonical untuned value from Cormack et al. 2009, fixed before inspecting outputs (Table 2 note). Not fitted.
Entity-index cleaning threshold for Aethel-NER3 = drop ~20k OCR fragments/bare numerals from ~27k raw extractions, leaving ~6k entities
A hand-chosen post-processing step. The paper discloses it and shows the regex variant (Aethel-Reg) fails without it, but the exact filtering criteria are not fully specified, so independent replication would require tuning.

axioms (4)

domain assumption Personalized PageRank over a bipartite entity-passage graph with column-normalized adjacency converges and ranks relevant passages via random-walk diffusion.
Section 4.3 relies on standard PPR convergence; standard math.
domain assumption The 20 multi-hop QA questions and 40 financial queries are representative of multi-hop retrieval workloads.
The core comparison is over a small manually constructed sample; the paper acknowledges this in Limitations.
domain assumption Gold passage labels for the financial corpus are correct despite single-annotator labeling.
Used to compute HR@5 for all systems in Table 2.
domain assumption Aethel-NER3's substring-adjacency entity-index construction is free of the failure mode that sank the naive-spaCy variant.
The main NER3 result depends on this corrected entity index; the paper explicitly discloses this in Section 6.3 and Limitations.

invented entities (2)

Bipartite Coreference Teleportation (BCT) layer independent evidence
purpose: Expands the PPR personalization vector via aliases and substring overlap to improve hit-rate coverage.
Its effect is measured against external benchmarks (MuSiQue/2WikiMultiHopQA HR@5) rather than asserted, so it has a falsifiable handle.
Specialist swarm (Liquidity Specialist, Valuation Auditor, Diligence Auditor, Market Research Specialist, Synthesizer) no independent evidence
purpose: Downstream synthesis of the retrieved evidence into an investment committee memo.
Explicitly stated to be not separately benchmarked in this paper (Section 3); system description only, so no independent evidence is provided here.

pith-pipeline@v1.3.0-alltime-deepseek · 10445 in / 6968 out tokens · 59990 ms · 2026-08-01T15:06:04.985131+00:00 · methodology

0 comments

read the original abstract

Secondary private equity transactions require rapid synthesis of fragmented, unstructured financial disclosures, where critical metrics and their entity anchors are distributed across disjoint documents with limited lexical overlap. We present Aethel, a reproducible framework that combines bipartite Personalized PageRank graph retrieval with a coreference-aware Bipartite Coreference Teleportation layer and an orchestrated specialist-agent architecture. Aethel models corpora as entity-passage graphs and propagates relevance through explicit relational paths to support multi-hop financial diligence. We evaluate the retrieval layer on 200-question samples from the MuSiQue and 2WikiMultiHopQA validation sets, comparing sparse lexical, dense bi-encoder, vanilla graph, and coreference-aware graph retrieval. Aethel achieves HR@5 of 100.0% on 2WikiMultiHopQA and 88.5% on MuSiQue, improving coverage over vanilla graph retrieval while trading off top-rank precision. We also evaluate retrieval over a 4,123-chunk corpus of financial disclosures and find that graph retrieval outperforms dense retrieval on multi-hop recall but does not surpass a strong BM25 baseline at open-corpus scale. The results show that graph-based retrieval offers interpretable multi-hop evidence paths and degrades more gracefully than dense retrieval as corpus size grows, while also demonstrating that its advantage depends strongly on corpus scale and entity-index quality. Code and evaluation artifacts are released for reproducibility.

Figures

Figures reproduced from arXiv: 2607.24826 by Krish Sapru.

**Figure 2.** Figure 2: Bipartite entity–passage retrieval in Aethel. Query entities seed the PPR personalization vector s; power iteration diffuses probability mass through entity–passage edges (no passage– passage edges exist). The red path illustrates a two-hop chain: seed entity e1 activates passage p1, which shares entity e3, propagating mass onward to passage p3. All passages are ranked by their steady-state probability p … view at source ↗

**Figure 3.** Figure 3: Multi-hop HR@5 as corpus size grows from 100 to 4,123 chunks, over the 20 multi-hop queries. Gold passages are retained in every subsample (see text). Bands show ±1σ over 5 resamples (seed 42); the full-corpus point is a single deterministic trial. Dense retrieval leads at the smallest pool (0.840 at 100 chunks) but collapses to 0.450 at full scale, while BM25 (0.800→0.700) is robust to distractor growt… view at source ↗

discussion (0)

Reference graph

Works this paper leans on

12 extracted references · 3 linked inside Pith

[1]

Darren Edge, Ha Trinh, Newman Cheng, Joshua Bradley, Alex Chao, Apurva Mody, Steven Truitt, Dasha Metropolitansky, Robert Osazuwa Ness, and Jonathan Larson. 2024. From local to global: A graph rag approach to query-focused summarization. arXiv preprint arXiv:2404.16130

Pith/arXiv arXiv 2024
[2]

Bernal Jim \'e nez Guti \'e rrez, Yiheng Shu, Yu Gu, Michihiro Yasunaga, and Yu Su. 2024. Hippo RAG : Neurobiologically inspired long-term memory for large language models. Advances in Neural Information Processing Systems (NeurIPS 2024)

2024
[3]

Bernal Jim \'e nez Guti \'e rrez, Yiheng Shu, Weijian Qi, Sizhe Zhou, and Yu Su. 2025. From RAG to memory: Non-parametric continual learning for large language models. Proceedings of the 42nd International Conference on Machine Learning (ICML 2025)

2025
[4]

Xanh Ho, Anh-Khoa Duong Nguyen, Saku Sugawara, and Akiko Aizawa. 2020. Constructing a multi-hop QA dataset for comprehensive evaluation of reasoning steps. Proceedings of the 28th International Conference on Computational Linguistics (COLING 2020), pages 6609--6625

2020
[5]

Vladimir Karpukhin, Barlas O g uz, Sewon Min, Patrick Lewis, Ledell Wu, Sergey Edunov, Danqi Chen, and Wen-tau Yih. 2020. Dense passage retrieval for open-domain question answering. Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing (EMNLP 2020), pages 6769--6781

2020
[6]

u ttler, Mike Lewis, Wen-tau Yih, Tim Rockt \

Patrick Lewis, Ethan Perez, Aleksandra Piktus, Fabio Petroni, Vladimir Karpukhin, Naman Goyal, Heinrich K \"u ttler, Mike Lewis, Wen-tau Yih, Tim Rockt \"a schel, Sebastian Riedel, and Douwe Kiela. 2020. Retrieval-augmented generation for knowledge-intensive nlp tasks. Advances in Neural Information Processing Systems (NeurIPS 2020)

2020
[7]

Lawrence Page, Sergey Brin, Rajeev Motwani, and Terry Winograd. 1999. The PageRank citation ranking: Bringing order to the web. Stanford InfoLab Technical Report

1999
[8]

Boci Peng, Yun Zhu, Yongchao Liu, Xiaohe Bo, Haizhou Shi, Chuntao Hong, Yan Zhang, and Siliang Tang. 2024. Graph retrieval-augmented generation: A survey. arXiv preprint arXiv:2408.08921

Pith/arXiv arXiv 2024
[9]

Bhaskarjit Sarmah, Benika Hall, Rohan Rao, Sunil Patel, Stefano Pasquali, and Dhagash Mehta. 2024. HybridRAG: Integrating knowledge graphs and vector retrieval augmented generation for efficient information extraction. Proceedings of the 5th ACM International Conference on AI in Finance (ICAIF 2024), pages 608--616

2024
[10]

Harsh Trivedi, Niranjan Balasubramanian, Tushar Khot, and Ashish Sabharwal. 2022. MuSiQue: Multihop Questions via Single Hop Question Composition. Transactions of the Association for Computational Linguistics (TACL), 10:539--554

2022
[11]

Cormack, Charles L

Gordon V. Cormack, Charles L. A. Clarke, and Stefan Buettcher. 2009. Reciprocal rank fusion outperforms condorcet and individual rank learning methods. Proceedings of the 32nd International ACM SIGIR Conference on Research and Development in Information Retrieval (SIGIR 2009), pages 758--759

2009
[12]

White, Doug Burger, and Chi Wang

Qingyun Wu, Gagan Bansal, Jieyu Zhang, Yiran Wu, Beibin Li, Erkang Zhu, Li Jiang, Xiaoyun Zhang, Shaokun Zhang, Jiale Liu, Ahmed Hassan Awadallah, Ryen W. White, Doug Burger, and Chi Wang. 2023. Autogen: Enabling next-gen LLM applications via multi-agent conversation framework. arXiv preprint arXiv:2308.08155

Pith/arXiv arXiv 2023

[1] [1]

Darren Edge, Ha Trinh, Newman Cheng, Joshua Bradley, Alex Chao, Apurva Mody, Steven Truitt, Dasha Metropolitansky, Robert Osazuwa Ness, and Jonathan Larson. 2024. From local to global: A graph rag approach to query-focused summarization. arXiv preprint arXiv:2404.16130

Pith/arXiv arXiv 2024

[2] [2]

Bernal Jim \'e nez Guti \'e rrez, Yiheng Shu, Yu Gu, Michihiro Yasunaga, and Yu Su. 2024. Hippo RAG : Neurobiologically inspired long-term memory for large language models. Advances in Neural Information Processing Systems (NeurIPS 2024)

2024

[3] [3]

Bernal Jim \'e nez Guti \'e rrez, Yiheng Shu, Weijian Qi, Sizhe Zhou, and Yu Su. 2025. From RAG to memory: Non-parametric continual learning for large language models. Proceedings of the 42nd International Conference on Machine Learning (ICML 2025)

2025

[4] [4]

Xanh Ho, Anh-Khoa Duong Nguyen, Saku Sugawara, and Akiko Aizawa. 2020. Constructing a multi-hop QA dataset for comprehensive evaluation of reasoning steps. Proceedings of the 28th International Conference on Computational Linguistics (COLING 2020), pages 6609--6625

2020

[5] [5]

Vladimir Karpukhin, Barlas O g uz, Sewon Min, Patrick Lewis, Ledell Wu, Sergey Edunov, Danqi Chen, and Wen-tau Yih. 2020. Dense passage retrieval for open-domain question answering. Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing (EMNLP 2020), pages 6769--6781

2020

[6] [6]

u ttler, Mike Lewis, Wen-tau Yih, Tim Rockt \

Patrick Lewis, Ethan Perez, Aleksandra Piktus, Fabio Petroni, Vladimir Karpukhin, Naman Goyal, Heinrich K \"u ttler, Mike Lewis, Wen-tau Yih, Tim Rockt \"a schel, Sebastian Riedel, and Douwe Kiela. 2020. Retrieval-augmented generation for knowledge-intensive nlp tasks. Advances in Neural Information Processing Systems (NeurIPS 2020)

2020

[7] [7]

Lawrence Page, Sergey Brin, Rajeev Motwani, and Terry Winograd. 1999. The PageRank citation ranking: Bringing order to the web. Stanford InfoLab Technical Report

1999

[8] [8]

Boci Peng, Yun Zhu, Yongchao Liu, Xiaohe Bo, Haizhou Shi, Chuntao Hong, Yan Zhang, and Siliang Tang. 2024. Graph retrieval-augmented generation: A survey. arXiv preprint arXiv:2408.08921

Pith/arXiv arXiv 2024

[9] [9]

Bhaskarjit Sarmah, Benika Hall, Rohan Rao, Sunil Patel, Stefano Pasquali, and Dhagash Mehta. 2024. HybridRAG: Integrating knowledge graphs and vector retrieval augmented generation for efficient information extraction. Proceedings of the 5th ACM International Conference on AI in Finance (ICAIF 2024), pages 608--616

2024

[10] [10]

Harsh Trivedi, Niranjan Balasubramanian, Tushar Khot, and Ashish Sabharwal. 2022. MuSiQue: Multihop Questions via Single Hop Question Composition. Transactions of the Association for Computational Linguistics (TACL), 10:539--554

2022

[11] [11]

Cormack, Charles L

Gordon V. Cormack, Charles L. A. Clarke, and Stefan Buettcher. 2009. Reciprocal rank fusion outperforms condorcet and individual rank learning methods. Proceedings of the 32nd International ACM SIGIR Conference on Research and Development in Information Retrieval (SIGIR 2009), pages 758--759

2009

[12] [12]

White, Doug Burger, and Chi Wang

Qingyun Wu, Gagan Bansal, Jieyu Zhang, Yiran Wu, Beibin Li, Erkang Zhu, Li Jiang, Xiaoyun Zhang, Shaokun Zhang, Jiale Liu, Ahmed Hassan Awadallah, Ryen W. White, Doug Burger, and Chi Wang. 2023. Autogen: Enabling next-gen LLM applications via multi-agent conversation framework. arXiv preprint arXiv:2308.08155

Pith/arXiv arXiv 2023