acceptodds
Under review as a conference paper at ICLR 2027

SAG: SQL-Retrieval Augmented Generation with Query-Time Dynamic Hyperedges

Abstract

While retrieval-augmented generation (RAG) has proven effective at giving LLMs access to external knowledge, mainstream dense-retrieval implementations remain inherently limited in handling structured constraints and multi-hop reasoning. Graph-based methods address this by constructing knowledge graphs offline, but they often fragment semantics, incur high maintenance, and complicate incremental updates. We propose SAG (SQL-Retrieval Augmented Generation), a structured retrieval architecture that organizes documents into an event-entity index without building a global knowledge graph. SAG represents each chunk as a semantically complete event paired with its entities, forming a latent hyperedge that preserves n-ary relations without decomposing them into triples. At query time, SAG treats shared entities as join keys to connect related chunks. This dynamically yields a query-scoped neighborhood of events, and yet every piece of evidence remains the original chunk throughout. Experiments on HotpotQA, 2WikiMultiHopQA, and MuSiQue show that SAG achieves the best retrieval and end-to-end QA performance on all three benchmarks. On MuSiQue, where multi-hop evidence chaining is most demanding, SAG reaches 80.36% Recall@5, outperforming the strongest baseline by 11.52 points. This work paves the way for knowledge infrastructure that enables LLM agents to retrieve and reason over continually growing organizational knowledge.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
RAG, Structured Retrieval, Multi-Hop Retrieval, Hyperedges, Knowledge Infrastructure

Reviews

Among:
Shape:
1234this paper, Top 32%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.50 over 4 reviews, above 68% of the others.
Reviewer_YkW4OpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes SAG, a SQL-augmented retrieval pipeline that leverages query-time dynamic hyperedges to improve multi-hop retrieval. The pipeline consists of two main stages:

  1. Offline indexing: An LLM summarizes each chunk into an event and extracts the associated entities. Events and entities are linked through many-to-many relationships stored in a SQL database.

  2. Online retrieval: SAG combines structured retrieval with query-based vector retrieval. Specifically, an LLM first extracts entities from the query, which are then used to retrieve associated events through SQL joins. In parallel, the query is embedded to retrieve semantically relevant events. The retrieved events are further expanded through shared entities, and the resulting candidate pool is ranked by query-event similarity to retain the top 100 events. An LLM then selects five events, which are mapped back to their original chunks. Finally, direct vector retrieval supplements these selected chunks with five additional, non-overlapping chunks, yielding ten chunks for answer generation.

Experiments on several multi-hop QA benchmarks demonstrate that SAG consistently outperforms the evaluated baselines.

Critical strengths
  1. The paper is well written, clearly organized, and easy to follow.

  2. The proposed pipeline represents a reasonable engineering approach to combining structured retrieval and dense retrieval.

Critical weaknesses
  1. Missing references. Several relevant studies are not adequately discussed. In particular, [1, 2] also propose structure-augmented RAG approaches and should be included in the related work to better position the contributions of SAG.

  2. Concerns about comparison fairness. The experimental comparisons do not adequately control for differences in candidate retrieval and reranking budgets. SAG retrieves and expands a relatively large candidate pool before employing a powerful LLM to select the final evidence, whereas the dense retrieval baselines do not appear to benefit from an equivalent selection process. Consequently, part of the reported improvements may be attributable to the larger candidate pool and stronger reranking rather than the proposed structural representation. This makes it difficult to isolate the effectiveness of the core method.

  3. Limited methodological novelty. The proposed method appears to be primarily an engineering integration of existing techniques, including entity-based indexing, structured retrieval, graph-style expansion through shared entities, and LLM-based evidence selection. While using SQL to implement these operations may offer practical advantages, the methodological novelty beyond existing structure-augmented retrieval approaches is not sufficiently established.

[1] LinearRAG: Linear Graph Retrieval Augmented Generation on Large-scale Corpora, Zhuang et. al, ICLR 2026.

[2] E$^2$GraphRAG: Streamlining Graph-based RAG for High Efficiency and Effectiveness, Zhao et.al, Arxiv preprint.

Questions for authors
  1. Comparison with stronger dense retrieval baselines. What happens if the dense retrieval baselines are allowed to retrieve a larger candidate pool (e.g., top 100) and employ the same LLM-based evidence selector as SAG? Would SAG still maintain its performance advantage under this setting?

  2. Analysis of structural retrieval gains. Could the authors provide a detailed analysis of how many gold supporting passages are recovered by SAG's structural expansion but are absent from the top 100 candidates retrieved by dense retrieval? Such an analysis would help distinguish improvements in candidate coverage from those attributable to LLM-based reranking.

2 comments
AuthorsOpenReview ↗

Response to Reviewer YkW4 (1/2)

We thank the reviewer for finding the paper well written and the pipeline a reasonable combination of structured and dense retrieval. The new references and baselines are in the revised PDF; the other experiments are reported for this discussion.


W1. LinearRAG and E²GraphRAG

Both are now discussed in Related Work and evaluated as baselines in Table 1, Table 3, and the appendix, together with HGRAG, BrowseNet, and SiReRAG. All methods use the same corpora, encoder (BGE), and reader. E²GraphRAG and SiReRAG return summary nodes or unordered sets, so we report QA for them.

MethodMuSiQue R@1/2/5/10MuSiQue EM/F12Wiki R@1/2/5/102Wiki EM/F1HotpotQA R@1/2/5/10HotpotQA EM/F1
SiReRAG–44.30 / 55.26–68.40 / 75.87–63.90 / 77.75
E²GraphRAG–16.80 / 24.09–41.70 / 46.23–48.50 / 59.28
HGRAG27.08 / 42.94 / 58.98 / 68.5637.30 / 45.6941.23 / 72.45 / 90.42 / 95.7568.70 / 77.3040.30 / 66.40 / 86.40 / 94.3060.50 / 72.37
LinearRAG20.78 / 31.58 / 49.14 / 61.4629.00 / 38.0138.85 / 65.58 / 84.88 / 94.3564.60 / 71.9836.20 / 58.95 / 77.85 / 89.2556.00 / 68.42
BrowseNet32.28 / 46.47 / 60.61 / 63.3934.60 / 44.4742.50 / 72.85 / 87.38 / 90.3568.90 / 76.0945.25 / 78.65 / 91.60 / 92.2562.70 / 75.36
SAG36.82 / 63.62 / 80.36 / 83.3750.20 / 61.1543.81 / 83.93 / 93.34 / 93.5770.50 / 78.0647.80 / 91.55 / 96.50 / 97.7067.90 / 79.66
  • SAG has the highest R@1, R@2, R@5, and F1 on all three datasets. Over LinearRAG, SAG gains 31.22 / 8.46 / 18.65 R@5 on MuSiQue / 2Wiki / HotpotQA.
  • Positioning. LinearRAG builds a hierarchical graph from co-occurrence without LLM extraction and ranks passages by global importance aggregation; E²GraphRAG pairs a summary tree with a noun co-occurrence graph for long-document QA. Both precompute their structure over the corpus, whereas SAG stores per-event incidence rows and assembles the cross-document structure for each query.

W2 + Q1 + Q2. Candidate-pool and selection budgets; gold passages beyond dense top-100

"What happens if the dense retrieval baselines are allowed to retrieve a larger candidate pool (e.g., top 100) and employ the same LLM-based evidence selector as SAG?" … "How many gold supporting passages are recovered by SAG's structural expansion but are absent from the top 100 candidates retrieved by dense retrieval?"

SAG maintains its advantage over dense retrieval with the candidate pool and selector equalized, and is on par with HippoRAG 2. On MuSiQue (1,000 queries), every system supplies 100 candidates, and the same Qwen3.6-Flash selector, prompt, and truncation choose 5:

Final evidence: 5 passages chosen by the selector.

Method (same LLM selector)R@1R@2R@5
BGE top-10035.8761.1275.11
SAG36.8263.6280.36
HippoRAG 2 top-10036.7065.2280.72

Candidate coverage: measured on each method's ranked candidates before the selector is applied.

CandidatesRecall@50Recall@100Gold passages in top-100
BGE81.0283.912,196 / 2,648 (82.93%)
SAG85.9989.322,334 / 2,648 (88.14%)
HippoRAG 287.2791.162,358 / 2,648 (89.05%)
  • Against dense retrieval with the same encoder (BGE) and the same selector, SAG raises R@5 by 5.25 and R@2 by 2.50. Its 100 candidates contain 138 more gold passages (2,334 vs. 2,196 of 2,648), and candidate Recall@100 rises from 83.91 to 89.32. With pool size and selector held fixed, the gain comes from the candidates SAG assembles.
  • Against HippoRAG 2 with the same selector, SAG is on par (R@5 80.36 vs. 80.72; R@1 36.82 vs. 36.70). HippoRAG 2 already uses an LLM to filter retrieved triples before choosing its seed nodes, so this configuration applies an LLM selector twice. SAG applies the LLM selector once and reaches the same recall without corpus-level PageRank propagation or synonym edges.

Gold passages beyond dense top-100 (Q2). With the selector held fixed, candidate coverage and final recall move together, so the gain comes from candidate coverage rather than from reranking. The additional gold passages are reached by hop-by-hop expansion through shared entities, which dense similarity to the question does not rank highly.

The ranking-and-pruning design is what separates SAG from methods that expand from their seeds in one pass over the corpus graph, and its effect is largest on evidence deep in the reasoning chain (see W3 in the next comment).

AuthorsOpenReview ↗

Response to Reviewer YkW4 (2/2)

W3. Methodological novelty

"The proposed method appears to be primarily an engineering integration of existing techniques … the methodological novelty beyond existing structure-augmented retrieval approaches is not sufficiently established."

SAG's contribution is how entity structure is used in retrieval, a design that differs from the closest graph methods, including LinearRAG named by the reviewer and HGRAG, the closest hypergraph design. The individual components are standard building blocks; the design that combines them is what produces the gains below.

1. Structure decides which candidates exist; it never enters the ranking score. LinearRAG and HGRAG propagate scores from the query entities over an entity graph (Personalized PageRank, hypergraph diffusion) and rank every passage by one global score, so a passage's rank depends on how much score reaches it from the query entities. In SAG, structure has a single job: a bounded one-hop expansion from the events already retrieved, along shared entities, with its size fixed by KseedK_{\text{seed}}, KfrontK_{\text{front}}, and KhopK_{\text{hop}}. Candidates are then ordered by query–event similarity alone, so the number of hops between a passage and the query entities does not affect its rank.

2. This design directly targets limitations the two papers state themselves.

  • Expansion growth. LinearRAG notes "the exponential growth of the search space as propagation deepens" (§3.2.1). SAG expands one hop with fixed budgets: on MuSiQue, LinearRAG activates 666 entities per query and adds entity scores to a median of 10,595 of the 11,656 passages, while SAG ranks at most 100 candidate passages.
  • Drift. LinearRAG notes that "irrelevant entities may repeatedly serve as new seeds … drift into semantic regions unrelated to the original query intent" (§3.2.1). In SAG, expanded entities never become new seeds, frontier entities are chosen by the relevance of their role descriptions to the query, and all candidates are re-ranked by similarity to the query.
  • Long reasoning chains. HGRAG notes that on datasets "such as MuSiQue requiring longer reasoning hops, structural modeling is more prone to incompleteness and errors." SAG's ranking contains no propagated score and no diffusion-step parameter, and one default configuration serves all three datasets, whereas HGRAG tunes its diffusion steps and weights per dataset.

3. The effect: deep evidence is ranked higher. We replayed both methods' scoring with their released code and grouped gold passages by their graph distance from the query's seed entities. For passages three or more steps away, the similarity term inside each method places 30.1% (HGRAG) and 39.0% (LinearRAG) in the top 5; after propagation, only 2.3% and 6.7% remain there. The same passages reach 35.4% and 38.1% in SAG's candidate order, because their distance does not change their rank.

4. Result on multi-hop evidence (MuSiQue; SAG's candidate order before the LLM selector, versus each method's final ranking):

Evidence (MuSiQue)PassagesMethodTop-5Top-20Top-50
Depth 3 (requires answering two prior hops)409SAG40.363.875.3
LinearRAG22.546.265.8
HGRAG20.040.659.4
Last hop of 3/4-hop questions (holds the answer)482SAG39.866.479.3
LinearRAG24.950.270.1
HGRAG26.146.162.9

The 420 gold passages that SAG ranks below position 20 all enter its candidates through the entity path (116) or the one-hop expansion (304), which is where the structure contributes.

The other two design properties support this contribution: event-level units preserve the n-ary relation of each chunk (replacing them with triples lowers R@5 by 2.75 / 9.24 / 13.80 and F1 by 9.95 / 3.65 / 6.80 on MuSiQue / 2Wiki / HotpotQA, Table 2), and the index is append-only, so new documents add rows without changing existing records.

Protocol. All methods use BGE and the same corpora, with the hyperparameters from their papers. With its threshold raised to 0.4, LinearRAG reaches 55.25 R@5 on a 200-query subset. HGRAG's reported 74.1 R@5 on MuSiQue uses NV-Embed-v2 and Llama-3.3-70B; under this shared protocol it reaches 58.98.


We thank the reviewer again. The revised paper adds the requested references and baselines. This response adds a controlled comparison with equal candidate pool and selector, and shows that SAG's contribution is a design in which structure assembles candidates while ranking follows relevance to the query.

Reviewer_X9YYOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes SAG, a RAG method that combines vector retrieval with entity-based connections. It extracts an event and related entities from each text chunk, then uses SQL joins to connect chunks that share entities. This helps find evidence across different documents while keeping the original text for answering questions. Experiments on three multi-hop QA datasets show improvements in both retrieval and answer quality.

Critical strengths
  1. SAG is a simple and practical design that use event-entity records and SQL to connect information across documents.
  2. SAG's experiment results show clear improvements on three multi-hop datasets.
Critical weaknesses
  1. The event–entity design is interesting, but its advantages over existing graph and hypergraph retrieval methods are not fully established. A more direct comparison with the closest methods would help.
  2. SAG uses an LLM to select evidence from the retrieved candidates. Comparing it with dense retrieval followed by the same LLM selector would help show how much improvement comes from the proposed structural retrieval.
  3. The experiments use relatively small corpora, however, the scalability experiments about costs is limited.
Questions for authors

Please strengthen your experiments as in my proposed weakness will make your work more convincing.

2 comments
AuthorsOpenReview ↗

Response to Reviewer X9YY (1/2)

We thank the reviewer for the positive assessment of SAG as a simple and practical design. We report the three requested experiments below; the new baselines are part of the revised PDF.


W1. Direct comparison with the closest graph and hypergraph methods

We added five recent structure-augmented methods to Table 1, Table 3, and the appendix: HGRAG (passages as hyperedges over entities, the closest design to SAG), BrowseNet, LinearRAG, SiReRAG, and E²GraphRAG. All use the same corpora, encoder (BGE), and reader. SiReRAG and E²GraphRAG return summary nodes or unordered sets, so we report QA for them.

R@k: Recall@k (%).

MethodStructureMuSiQue R@1/2/5/10MuSiQue EM/F12Wiki R@1/2/5/102Wiki EM/F1HotpotQA R@1/2/5/10HotpotQA EM/F1
HGRAGhypergraph, passages as hyperedges27.08 / 42.94 / 58.98 / 68.5637.30 / 45.6941.23 / 72.45 / 90.42 / 95.7568.70 / 77.3040.30 / 66.40 / 86.40 / 94.3060.50 / 72.37
BrowseNetchunk graph + traversal32.28 / 46.47 / 60.61 / 63.3934.60 / 44.4742.50 / 72.85 / 87.38 / 90.3568.90 / 76.0945.25 / 78.65 / 91.60 / 92.2562.70 / 75.36
LinearRAGthree-layer co-occurrence graph20.78 / 31.58 / 49.14 / 61.4629.00 / 38.0138.85 / 65.58 / 84.88 / 94.3564.60 / 71.9836.20 / 58.95 / 77.85 / 89.2556.00 / 68.42
SiReRAGsimilarity / relatedness trees–44.30 / 55.26–68.40 / 75.87–63.90 / 77.75
E²GraphRAGsummary tree + noun graph–16.80 / 24.09–41.70 / 46.23–48.50 / 59.28
SAGevent–entity incidence rows36.82 / 63.62 / 80.36 / 83.3750.20 / 61.1543.81 / 83.93 / 93.34 / 93.5770.50 / 78.0647.80 / 91.55 / 96.50 / 97.7067.90 / 79.66
  • SAG has the highest Recall@1, Recall@2, Recall@5, and F1 on all three datasets.
  • Over HGRAG, SAG gains 21.38 / 2.92 / 10.10 Recall@5 on MuSiQue / 2WikiMultiHopQA / HotpotQA. The margin is largest on MuSiQue, where SAG leads the next-best method (BrowseNet) by 19.75 Recall@5.

Why SAG leads the graph and hypergraph methods. HGRAG and LinearRAG rank passages by scores propagated over the corpus graph; SAG uses its structure only to assemble a bounded candidate set and ranks candidates by similarity to the query. The difference shows on evidence deep in the reasoning chain (MuSiQue; SAG's candidate order before the LLM selector, versus each method's final ranking):

Evidence (MuSiQue)PassagesMethodTop-5Top-20Top-50
Depth 3 (requires answering two prior hops)409SAG40.363.875.3
LinearRAG22.546.265.8
HGRAG20.040.659.4
Last hop of 3/4-hop questions (holds the answer)482SAG39.866.479.3
LinearRAG24.950.270.1
HGRAG26.146.162.9
AuthorsOpenReview ↗

Response to Reviewer X9YY (2/2)

W2. Dense retrieval followed by the same LLM selector

With the same encoder and selector, SAG's structural retrieval adds 5.25 Recall@5 over dense retrieval. On MuSiQue (1,000 queries), each system supplies 100 candidates and the same Qwen3.6-Flash selector chooses up to 5:

Method (same LLM selector)R@5
BGE top-10075.11
SAG80.36
HippoRAG 2 top-10080.72

Candidate coverage, measured before the selector is applied:

CandidatesRecall@100Gold passages in top-100
BGE83.912,196 / 2,648
SAG89.322,334 / 2,648
HippoRAG 291.162,358 / 2,648
  • SAG's candidates contain 138 more gold passages than dense top-100; with the selector fixed, this is the contribution of structural retrieval.
  • SAG is on par with HippoRAG 2 (80.36 vs. 80.72 Recall@5) while using the LLM selector once; HippoRAG 2 also uses an LLM to filter triples when choosing its seed nodes.

W3. Corpus size and cost scalability

Full-scale open-domain 2WikiMultiHopQA. The main protocol evaluates 1,000 questions sampled from the 2WikiMultiHopQA development set, with a corpus built from their context paragraphs (6,119 passages). We now run the full setting: all 12,576 development questions of 2WikiMultiHopQA [1] over the open-domain corpus introduced by IRCoT [2], which pools and deduplicates the context paragraphs of all training, development, and test questions into 430,225 Wikipedia passages, 70× the main-protocol corpus. The questions carry 30,654 gold supporting passages (2.44 per question) across compositional (41.6%), comparison (24.2%), bridge-comparison (21.9%), and inference (12.3%) questions. Running the graph baselines at this scale is the main cost of the experiment: building HippoRAG 2's graph over 430K passages requires more than 50 GB of memory, and we secured these resources for this response. All methods use BGE and the same reader; indexing runs with 10 concurrent LLM calls for every method, and retrieval time is measured per query at concurrency 1.

MethodR@1R@2R@5R@10EMF1
BGE32.6548.2957.3461.0144.9750.30
HippoRAG 234.7655.2666.5070.0952.5858.55
HGRAG30.3649.1668.2777.5056.1661.49
SAG36.8766.4080.8082.3165.7772.67

Cost in the format of Table 3 of the paper: offline values are corpus totals; online values are per-query averages.

MethodF1Offline time (h)Offline tokens (M)Online retrieval (s)Online tokens
BGE50.300.5300.221,039
HippoRAG 258.551775795.044,882
HGRAG61.4913167.0521.881,330
SAG72.6779.51,88015.7212,691
  • Accuracy. SAG leads on every metric, with 12.53 Recall@5 and 11.18 F1 over HGRAG. Its lead over HGRAG grows from 2.92 Recall@5 on the 6,119-passage corpus to 12.53 on the 430K-passage corpus.
  • Cost. At the same indexing concurrency, SAG builds the 430K-passage index in 79.5 h, less than half of HippoRAG 2's 177 h, and its per-query retrieval (15.72 s with the paper's selector) is faster than HGRAG's (21.88 s).

We will add these full-scale results, extended to more baselines, to the final version of the paper.

Scaling evidence in the paper. With KpoolK_{\text{pool}} and KcandK_{\text{cand}} fixed, the per-query SQL and selection work does not grow with the corpus (Figure 3b), and SAG's Recall@5 changes less than HippoRAG 2's as the corpus grows (Figure 3c).

[1] Xanh Ho, Anh-Khoa Duong Nguyen, Saku Sugawara, Akiko Aizawa. Constructing A Multi-hop QA Dataset for Comprehensive Evaluation of Reasoning Steps. COLING 2020. [2] Harsh Trivedi, Niranjan Balasubramanian, Tushar Khot, Ashish Sabharwal. Interleaving Retrieval with Chain-of-Thought Reasoning for Knowledge-Intensive Multi-Step Questions. ACL 2023.


We thank the reviewer again for the concrete suggestions, which this response answers with direct comparisons to the closest methods and a controlled comparison under the same selector.

Reviewer_byEdOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper tries to propose a new RAG paradigm by transforming each text chunk into an event and corresponding to multiple entities. Then, a graph is constructed as the retrieval corpus. The framework is built upon offline and online stages, and experiments are conducted on various QA datasets. The performance demonstrates the effectiveness of the proposed approach, however, the experimental setting and the original claim might not be fully convincing.

Critical strengths
  1. The original motivation is good, while the vanilla RAG and graph-based RAG focus on different granularities, and the proposed SAG attempts to conduct RAG at a different level, which could address some RAG issues in the literature.

  2. The designed framework is clear and easy to follow, which shows a feasible way to build an event-based graph.

  3. The experimental results demonstrate some of the effectiveness of the proposed framework, and I believe such a design could bring some improvement in terms of designing a RAG system.

Critical weaknesses

There are some points that need to be polished and improved in terms of the current manuscript.

  1. I cannot see any differences between SAG and the existing text+knowledge graph RAG. The previous literature aims to retrieve knowledge from unstructured chunks and a structured graph at different granularities, so as to supply each other. However, the SAG does not distinguish the differences compared to them, and it would be better to investigate the combination of SAG and vanilla RAG.

  2. The experimental setting is not convincing, as it only uses 1000 queries for each dataset, and the retrieval is not conducted under an open setting, e.g., DPR (Dense Passage Retrieval for Open-Domain Question Answering). Containing only thousands of passages in the corpus is not feasible.

  3. The introduction claims that vanilla RAG is good at speed and cost, while SAG is fair. However, as shown in Table 3, the efficiency of SAG is so bad, as it requires 159× compared to Dense RAG at inference time. This is infeasible to deploy in common practice.

Questions for authors
  1. Please indicate the differences compared to existing text+knowledge graph RAG and show to what aspect the improvements are brought by SAG.

  2. Please using correct experimental setting to conduct experiments, i.e., full test set and the whole wikipedia corpus (around thousand hundreds of passages) following the previsou studies (selfRAG, search-r1).

  3. Please demonstrate the feasibility of deploying such a system or optimization for efficiency at inference time.

3 comments
AuthorsOpenReview ↗

Response to Reviewer byEd (1/3)

We thank the reviewer for finding the event-level motivation sound and the framework clear. The revised PDF contains the paper changes (positioning in Sections 1–3, Figure 1, new baselines, ablation F1); the other analyses are reported for this discussion.


W1 + Q1. Difference from text + knowledge-graph RAG, and where the improvement comes from

"I cannot see any differences between SAG and the existing text+knowledge graph RAG … show to what aspect the improvements are brought by SAG."

Text + KG methods build a graph alongside the text and retrieve from both; SAG keeps one append-only relational index and builds the cross-document structure for each query. The revised paper (Sections 1, 2, 3.1) states this difference; the table below summarizes it:

MethodRetrieval unitCorpus-level structure computed offlineAt query time
GraphRAGentities, relations, communitiescommunity summaries; one merged description per entitycommunity / entity retrieval
LightRAGentities, relationsone merged description per entitykeyword + graph retrieval
HippoRAG 2passages + entitiessynonym edges; graph for Personalized PageRankpropagation over the corpus graph
HGRAGpassages as hyperedgesglobal incidence matrix and Laplacianhypergraph diffusion
SAGone event per chunk + typed entitiesnone; append-only incidence rowsbounded SQL joins assemble the neighborhood

Each incidence row stores a role description of what the entity does in that event, and descriptions are never merged across events. A new document therefore appends rows without changing any existing record, and expansion scores an entity by its role in the seed event rather than by how often it occurs.

Where the improvement comes from: structure decides which candidates exist, not how they are ranked. Text + KG methods use their graph to compute the ranking score: relevance propagates from the query entities over the corpus graph, so evidence several reasoning steps away receives little score. SAG uses its structure only to assemble candidates through a bounded one-hop expansion along shared entities, and ranks them by similarity to the query. Replaying the scoring of HGRAG and LinearRAG with their released code shows the effect: for gold passages three or more steps from the query's seed entities, the similarity term inside each method places 30.1% and 39.0% in the top 5, but after propagation only 2.3% and 6.7% remain; the same passages reach 35.4% and 38.1% in SAG's candidate order. The improvement is therefore concentrated on evidence deep in the reasoning chain (MuSiQue; SAG's candidate order before the LLM selector, versus each method's final ranking):

Evidence (MuSiQue)PassagesMethodTop-5Top-20Top-50
Depth 3 (requires answering two prior hops)409SAG40.363.875.3
LinearRAG22.546.265.8
HGRAG20.040.659.4
Last hop of 3/4-hop questions (holds the answer)482SAG39.866.479.3
LinearRAG24.950.270.1
HGRAG26.146.162.9

The 420 gold passages that SAG ranks below position 20 all enter its candidates through the entity path (116) or the one-hop expansion (304), which is where the structure contributes. On the main results, SAG's largest margin is on MuSiQue, the benchmark with the longest chains: 11.52 R@5 and 5.89 F1 over the strongest baseline (Table 1). The ablations (Table 2, now with QA F1) confirm each mechanism: replacing event indexing with triples, removing expansion, or replacing the selector lowers both Recall@5 and F1 on all three datasets.

Combination with vanilla RAG. SAG combines dense and structural retrieval at two levels. Inside structural retrieval, dense similarity sets the priority at every pruning step: seed events, frontier entities (through their role descriptions), and expanded events are each ranked by embedding similarity to the query before truncation. At the output, 5 chunks from structural selection are joined by 5 non-overlapping chunks from dense retrieval. The appendix sweeps this split on MuSiQue: dense only reaches 56.23 R@5, structural only 79.80, and the 5 + 5 combination 80.36. We agree that this combination merits further study and will explore tighter ways to combine the two paths.

AuthorsOpenReview ↗

Response to Reviewer byEd (2/3)

W2 + Q2. Evaluation setting

"Please use the full test set and the whole Wikipedia corpus following previous studies (Self-RAG, Search-R1)."

We now report full-scale open-domain results on 2WikiMultiHopQA. The main protocol evaluates 1,000 questions sampled from the 2WikiMultiHopQA development set, with a corpus built from their context paragraphs (6,119 passages). We now run the full setting: all 12,576 development questions of 2WikiMultiHopQA [1] over the open-domain corpus introduced by IRCoT [2], which pools and deduplicates the context paragraphs of all training, development, and test questions into 430,225 Wikipedia passages, 70× the main-protocol corpus. The questions carry 30,654 gold supporting passages (2.44 per question) across compositional (41.6%), comparison (24.2%), bridge-comparison (21.9%), and inference (12.3%) questions. Running the graph baselines at this scale is the main cost of the experiment: building HippoRAG 2's graph over 430K passages requires more than 50 GB of memory, and we secured these resources for this response. All methods use BGE and the same reader; indexing runs with 10 concurrent LLM calls for every method, and retrieval time is measured per query at concurrency 1.

MethodR@1R@2R@5R@10EMF1
BGE32.6548.2957.3461.0144.9750.30
HippoRAG 234.7655.2666.5070.0952.5858.55
HGRAG30.3649.1668.2777.5056.1661.49
SAG36.8766.4080.8082.3165.7772.67
  • SAG leads on every metric, with 12.53 R@5 and 11.18 F1 over the next-best method (HGRAG).
  • The margin grows with the corpus: on the 6,119-passage 2WikiMultiHopQA corpus of the main protocol, SAG leads HGRAG by 2.92 R@5 (93.34 vs. 90.42); on the 430K-passage corpus, by 12.53.
  • Per-query retrieval on the 430K-passage corpus takes 15.72 s with the paper's selector, faster than HGRAG (21.88 s).

This addresses the setting the reviewer points to: the full development set and an open-domain Wikipedia corpus on the scale of hundreds of thousands of passages. We will add these full-scale results, extended to more baselines, to the final version of the paper.

Main protocol. The main results follow the evaluation protocol released by HippoRAG 2, i.e., the same 1,000-question subsets and passage corpora, so that SAG is compared directly with published structure-augmented methods; HGRAG, BrowseNet, and SiReRAG report on this protocol as well.

Scaling evidence in the paper. With KpoolK_{\text{pool}} and KcandK_{\text{cand}} fixed, the per-query SQL and selection work does not grow with the corpus (Figure 3b), and SAG's R@5 changes less than HippoRAG 2's as the corpus grows (Figure 3c).

[1] Xanh Ho, Anh-Khoa Duong Nguyen, Saku Sugawara, Akiko Aizawa. Constructing A Multi-hop QA Dataset for Comprehensive Evaluation of Reasoning Steps. COLING 2020. [2] Harsh Trivedi, Niranjan Balasubramanian, Tushar Khot, Ashish Sabharwal. Interleaving Retrieval with Chain-of-Thought Reasoning for Knowledge-Intensive Multi-Step Questions. ACL 2023.

AuthorsOpenReview ↗

Response to Reviewer byEd (3/3)

W3 + Q3. Inference efficiency and deployment

"The efficiency of SAG is so bad, as it requires 159× compared to Dense RAG at inference time."

SAG's retrieval excluding the LLM selector takes 0.95 s per query, and the remaining time is spent in the LLM selector. The paper uses Qwen3.6-Flash, the open-weight Qwen3.6-35B-A3B, so that SAG can be reproduced and deployed with openly available weights. This model favors long reasoning before giving its answer [3]. Under the same prompt and the same candidates, the selector time therefore depends on the model: Qwen3.6-Flash takes 17.72 s, whereas JEV takes 0.34 s and the gpt-6 models 1.15–8.33 s. Per-stage times on MuSiQue (1,000 queries; Kcand=100K_{\text{cand}}=100 row of Table 3):

SelectorF1Retrieval excluding the LLM selector (s)LLM selector (s)Retrieval total (s)
Qwen3.6-Flash (paper setting)61.150.9517.7218.67
JEV65.480.950.341.29
gpt-6-luna62.491.038.339.36
gpt-6-sol65.331.075.916.98
gpt-6.1-sol-ultrafast63.280.991.152.14

Retrieval excluding the LLM selector: seed retrieval, expansion, and coarse ranking. LLM selector: the LLM choosing up to 5 events from the candidates.

In practice, we call the selector through an API: the JEV decision model through Typesafe's official API, or gpt-6-luna, gpt-6-sol, and gpt-6.1-sol-ultrafast. Switching the selector leaves the index and retrieval pipeline unchanged. With JEV, retrieval takes 1.29 s at 65.48 F1, on par with HippoRAG 2 (0.66 s); with gpt-6.1-sol-ultrafast, it takes 2.14 s, close to HGRAG (1.88 s) and BrowseNet (1.77 s). Across the gpt-6 models, retrieval takes 2.14–9.36 s at 62.49–65.33 F1, above the paper setting. For local deployment, we have recently served the open KnowLine-35B-A3B-Gen1.5 model as the selector, which reaches accuracy and speed similar to JEV.

  • Offline cost. SAG indexes MuSiQue in 4.57 h with 28.26M tokens, less than GraphRAG (10.15 h / 125.12M), LightRAG (25.94 h / 103.14M), HyperGraphRAG (16.21 h / 216.11M), and SiReRAG (7.50 h / 33.56M).

[3] Qwen Team. Qwen3.6-35B-A3B model card. https://huggingface.co/Qwen/Qwen3.6-35B-A3B


We thank the reviewer again. The revised paper states how SAG differs from text + KG RAG; this response further shows that SAG already combines dense and structural retrieval, and gives a feasible deployment configuration.

Reviewer_yjmsOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

SAG indexes each document chunk via an LLM (Qwen3.6-Flash) into one event statement and typed entities, storing them as relational rows in MySQL and vectors in Elasticsearch. At query time, it retrieves seed events via entity matching and dense search, expands them through 1-hop SQL joins over shared entities, and uses an LLM selector to pick 5 structural chunks alongside 5 dense-retrieved chunks. It achieves state-of-the-art Recall@5 (80.36% on MuSiQue) and QA F1 across HotpotQA, 2WikiMultiHopQA, MuSiQue, and NarrativeQA.

Critical strengths
  1. No Semantic Fragmentation: Uses event-entity pairs for SQL routing but returns original raw chunks for QA, avoiding the information loss of binary triples.
  2. Append-Only & Scalable: Independent chunk processing allows continual writes without graph reconstruction, and pre-filtering to Kpool=500K_{\text{pool}}=500 chunks retains ~99% of full-corpus Recall@5.
  3. Encoder Robustness: Recall@5 drops by only 2.46 points when swapping NV-Embed-v2 for BGE-v1.5 on MuSiQue, compared to 9.42 points for HippoRAG 2.
Critical weaknesses
  1. Confounded Retrieval Comparison: SAG's retrieval uses a Chain-of-Thought LLM selector (Qwen3.6-Flash) over 100 candidates, while baselines do not. Replacing the LLM selector with Qwen3-Reranker-8B drops Recall@5 by up to 15.29 points (falling below HippoRAG 2 on 2WikiMultiHopQA), whereas disabling SQL expansion (L=0L=0) still beats all baselines on MuSiQue (69.41%) and HotpotQA (95.30%). 2) Overclaimed "Non-Graph" Novelty: The SQL table is simply a materialized bipartite incidence graph, and the prompt example in Figure 6 copies the raw chunk verbatim as the "event," making it a standard chunk-entity graph.
  2. Underspecified Frontier Pruning: Section 4.2 sets seed and frontier budgets to 50, but Section 3.4 never specifies how entities are ranked or pruned to prevent high-frequency hubs (e.g., "United Kingdom") from flooding the SQL joins.
  3. High Online Latency: Despite Figure 1 claiming "Medium speed," SAG takes 25.03s per query—slower than both GraphRAG (12.79s) and HippoRAG 2 (9.16s).
  4. Missing baselines: Some of the latest baselines in multihop retrieval like BrowseNet and SireRAG are not included for comparison
Questions for authors
  1. Control for the LLM Selector: Evaluate BGE-v1.5 and HippoRAG 2 with the same Qwen3.6-Flash top-100 selector to isolate the true benefit of SQL expansion, and report QA F1 for Table 2 ablations.
  2. Formalize Pruning: State the exact scoring criteria used to prune the seed set (Kseed=50K_{\text{seed}}=50) and entity frontier (50) in Section 3.4.
  3. Add Alias Joins: Apply the entity vector threshold (τent=0.9\tau_{\text{ent}} = 0.9) to link synonymous entities during SQL expansion, addressing the alias limitation noted in Section 5.3.
  4. Calibrate Claims: Reframe the method around append-only relational indexing rather than the absence of a global graph, fix the verbatim copy in Figure 6, and align Figure 1's speed labels with Table 3
5 comments
AuthorsOpenReview ↗

Response to Reviewer yjms (1/3)

We thank the reviewer for the detailed review and for noting that SAG returns original chunks and supports append-only writes without graph reconstruction. The revised PDF contains the requested paper changes (framing, Sections 3.3–3.4, Figures 1 and 6, baselines, ablation F1); the other experiments are reported for this discussion.


W1 + Q1. Controlling for the LLM selector, and QA F1 for the ablations

"Evaluate BGE-v1.5 and HippoRAG 2 with the same Qwen3.6-Flash top-100 selector to isolate the true benefit of SQL expansion, and report QA F1 for Table 2 ablations."

Under the same selector and candidate budget, SAG outperforms dense retrieval and matches HippoRAG 2. We ran the requested control on MuSiQue (1,000 queries): each system supplies 100 candidates, and the same Qwen3.6-Flash selector, prompt, and truncation choose 5.

Final evidence: 5 passages chosen by the selector.

Method (same LLM selector)R@1R@2R@5
BGE top-10035.8761.1275.11
SAG36.8263.6280.36
HippoRAG 2 top-10036.7065.2280.72

Candidate coverage: measured on each method's ranked candidates before the selector is applied.

CandidatesRecall@50Recall@100Gold passages in top-100
BGE81.0283.912,196 / 2,648 (82.93%)
SAG85.9989.322,334 / 2,648 (88.14%)
HippoRAG 287.2791.162,358 / 2,648 (89.05%)
  • Under the same selector, SAG's expansion recovers 138 more gold passages than dense retrieval and raises R@5 by 5.25; against HippoRAG 2, it reaches the same recall with an append-only index.
  • Against dense retrieval with the same encoder (BGE), candidate Recall@100 rises from 83.91 to 89.32 (2,334 vs. 2,196 of 2,648 gold passages), and R@2 rises by 2.50. With pool size and selector held fixed, the gain comes from the candidates SAG assembles.
  • Against HippoRAG 2 with the same selector, SAG is on par (R@5 80.36 vs. 80.72; R@1 36.82 vs. 36.70). HippoRAG 2 already uses an LLM to filter retrieved triples before choosing its seed nodes, so this configuration applies an LLM selector twice. SAG applies the LLM selector once and reaches the same recall without corpus-level PageRank propagation or synonym edges.
  • Why SAG's candidates are stronger. Each pruned hop adds evidence linked to the previous hop through a shared entity, so the selector receives evidence chains rather than independent passages (mechanism in W2, next comment).

QA F1 for the ablations is now in Table 2. Every variant lowers F1 on all three benchmarks, by 1.68–9.95 points, so the retrieval differences carry through to answer quality:

VariantMuSiQue F12Wiki F1HotpotQA F1
Full SAG61.1578.0679.66
w/ triple indexing51.2074.4172.86
w/o expansion (L=0L=0)55.5470.7177.98
w/ Qwen3-Reranker-8B54.0370.1277.00
AuthorsOpenReview ↗

Response to Reviewer yjms (2/3)

W5. Missing baselines

We added BrowseNet and SiReRAG, together with LinearRAG, HGRAG, and E²GraphRAG, to Table 1, Table 3, and the full retrieval and QA tables in the appendix. All methods use the same corpora, encoder (BGE), and reader. SiReRAG and E²GraphRAG return summary nodes or unordered sets rather than ranked passages, so we report QA for them.

MethodMuSiQue R@1/2/5/10MuSiQue EM/F12Wiki R@1/2/5/102Wiki EM/F1HotpotQA R@1/2/5/10HotpotQA EM/F1
SiReRAG–44.30 / 55.26–68.40 / 75.87–63.90 / 77.75
E²GraphRAG–16.80 / 24.09–41.70 / 46.23–48.50 / 59.28
HGRAG27.08 / 42.94 / 58.98 / 68.5637.30 / 45.6941.23 / 72.45 / 90.42 / 95.7568.70 / 77.3040.30 / 66.40 / 86.40 / 94.3060.50 / 72.37
LinearRAG20.78 / 31.58 / 49.14 / 61.4629.00 / 38.0138.85 / 65.58 / 84.88 / 94.3564.60 / 71.9836.20 / 58.95 / 77.85 / 89.2556.00 / 68.42
BrowseNet32.28 / 46.47 / 60.61 / 63.3934.60 / 44.4742.50 / 72.85 / 87.38 / 90.3568.90 / 76.0945.25 / 78.65 / 91.60 / 92.2562.70 / 75.36
SAG36.82 / 63.62 / 80.36 / 83.3750.20 / 61.1543.81 / 83.93 / 93.34 / 93.5770.50 / 78.0647.80 / 91.55 / 96.50 / 97.7067.90 / 79.66
  • SAG has the highest R@1, R@2, R@5, and F1 on all three datasets. On MuSiQue / 2Wiki / HotpotQA, SAG gains 19.75 / 5.96 / 4.90 R@5 over BrowseNet and 5.89 / 2.19 / 1.91 F1 over SiReRAG.

W2 + Q4. Framing around append-only relational indexing; Figure 6; Figure 1

"The SQL table is simply a materialized bipartite incidence graph … Reframe the method around append-only relational indexing rather than the absence of a global graph."

We have adopted this framing throughout (abstract, introduction, Sections 3.1–3.2, related work, conclusion, Appendix A). The event–entity table is the incidence graph (Levi graph) of the event hypergraph, a standard lossless encoding (Berge, 1973). SAG's contribution lies in what is computed over this structure:

  • GraphRAG builds community summaries, and HippoRAG 2 adds synonym edges and propagates relevance with Personalized PageRank over the corpus graph. GraphRAG and LightRAG also consolidate all mentions of an entity into one corpus-level description.
  • SAG computes no corpus-level structure. Each incidence row stores a role description, i.e., what the entity does in that specific event, and these descriptions are never merged across events. Ingesting a chunk appends rows and leaves existing records unchanged, and cross-document structure is assembled per query by bounded joins.

How SAG uses the structure differs from graph propagation. Graph methods such as HippoRAG 2, LinearRAG, and HGRAG propagate scores over the corpus graph and rank passages by the propagated score, so evidence far from the query entities is pushed down. In SAG, the structure only assembles a bounded candidate set by one-hop expansion along shared entities, and candidates are ranked by query–event similarity alone. Evidence that requires several reasoning steps is therefore ranked higher (MuSiQue; SAG's candidate order before the LLM selector, versus each method's final ranking):

Evidence (MuSiQue)PassagesMethodTop-5Top-20Top-50
Depth 3 (requires answering two prior hops)409SAG40.363.875.3
LinearRAG22.546.265.8
HGRAG20.040.659.4
Last hop of 3/4-hop questions (holds the answer)482SAG39.866.479.3
LinearRAG24.950.270.1
HGRAG26.146.162.9

Figure 6. The original figure used an unrepresentative example: a chunk whose event matches the source text verbatim. Such cases occur in 0.5% of the 11,656 MuSiQue events. Passages in MuSiQue, 2WikiMultiHopQA, and HotpotQA are short, information-dense Wikipedia paragraphs, so a faithful single-event summary occasionally reproduces the paragraph. Across the corpus, the median character similarity between an event and its chunk is 0.67: events are rewritten summaries. The revised Figure 6 shows a representative output from the index used in our experiments together with all four messages of the extraction call, including the fixed demonstration.

Figure 1. SAG's speed label now matches Table 3.

AuthorsOpenReview ↗

Response to Reviewer yjms (3/3)

W3 + Q2 + Q3. Ranking and pruning, hub entities, and alias joins

"Section 3.4 never specifies how entities are ranked or pruned to prevent high-frequency hubs (e.g., 'United Kingdom') from flooding the SQL joins."

Sections 3.3–3.4 now state the exact criteria. Ranking always precedes truncation:

  1. Seed events. Rank the initial events by query–event similarity and keep the top Kseed=50K_{\text{seed}}=50.
  2. Frontier entities. Score each entity by the role description of its incidence row in a seed event against the query, taking its best-matching row; take the top 3Kfront3K_{\text{front}} rows, drop rows below 0.45 and already visited entities, and keep Kfront=50K_{\text{front}}=50 entities.
  3. Expanded events. A SQL join retrieves every unvisited event incident to the frontier; these events are ranked by query–event similarity and the top Khop=50K_{\text{hop}}=50 with score ≥ 0.4 are added.
  4. Query entities (Path A). Each entity's incident events are ranked by query–event similarity; the top 10 with score ≥ 0.3 are kept.

Hub entities. A hub such as "United Kingdom" is incident to a large number of events, so a join on it returns many rows. SAG handles this with similarity-based pruning at two points. First, the hub enters the frontier only when its role in a seed event matches the query; how many other events mention it does not raise its score. Second, once it is in the frontier, all of its incident events are ranked by similarity to the query before truncation, and each round passes at most Khop=50K_{\text{hop}}=50 events downstream. The number of candidates handed to the selector is therefore fixed regardless of entity degree, and the events kept from a hub are its most query-relevant ones.

Alias joins (Q3). SAG already uses this mechanism: Path A applies the entity vector threshold τent=0.9\tau_{\text{ent}}=0.9 to match the seed entities in the query against the entities in the index (Section 3.3), so different surface forms of the same entity connect at seed retrieval. We will extend the same matching to links between entities in the index, and add an alias table that maps full names, abbreviations, and aliases to a shared join key. Both changes act only on the join key and keep the index append-only.


W4. Online latency

"SAG takes 25.03s per query — slower than both GraphRAG (12.79s) and HippoRAG 2 (9.16s)."

SAG's retrieval excluding the LLM selector takes 0.95 s per query, and the rest of the retrieval time is spent in the selector. In the paper, the selector is Qwen3.6-Flash. This setting is slow because the selector prompt asks the model to write out its analysis of the candidates in the response before naming the chosen events (the model's thinking mode is not enabled). Per-stage times on MuSiQue (1,000 queries; Kcand=100K_{\text{cand}}=100 row of Table 3):

SelectorF1Retrieval excluding the LLM selector (s)LLM selector (s)Retrieval total (s)
Qwen3.6-Flash (paper setting)61.150.9517.7218.67
JEV65.480.950.341.29
gpt-6-luna62.491.038.339.36
gpt-6-sol65.331.075.916.98
gpt-6.1-sol-ultrafast63.280.991.152.14

Retrieval excluding the LLM selector: seed retrieval, expansion, and coarse ranking. LLM selector: the LLM choosing 5 events from the candidates.

In practice, we call the selector through an API: the JEV decision model through Typesafe's official API, or gpt-6-luna, gpt-6-sol, and gpt-6.1-sol-ultrafast. Switching the selector leaves the index and retrieval pipeline unchanged. With JEV, retrieval takes 1.29 s at 65.48 F1, on par with HippoRAG 2 (0.66 s); with gpt-6.1-sol-ultrafast, it takes 2.14 s, close to HGRAG (1.88 s) and BrowseNet (1.77 s). Across the gpt-6 models, retrieval takes 2.14–9.36 s at 62.49–65.33 F1, above the paper setting. For local deployment, we have recently served the open KnowLine-35B-A3B-Gen1.5 model as the selector, which reaches accuracy and speed similar to JEV.

  • Offline, SAG indexes MuSiQue in 4.57 h with 28.26M tokens, less than GraphRAG (10.15 h / 125.12M), LightRAG (25.94 h / 103.14M), HyperGraphRAG (16.21 h / 216.11M), and SiReRAG (7.50 h / 33.56M).

We thank the reviewer again. The revised paper now frames SAG around its append-only relational index, specifies every ranking and pruning step, and adds five recent baselines.

Comments on the paper

1 comment

Dear Reviewers,

Please take a moment to read the authors’ rebuttal and participate in the discussion.

Thanks,

AC

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.