QUEST: Query Optimization in Unstructured Document Analysis

Chengliang Chai; Guoren Wang; Han Han; Kaisen Jin; Lei Cao; Qiyan Deng; Xinyu Guo; Ye Yuan; Zhaoze Sun

REVIEW 5 major objections 5 minor 46 references

QUEST: Query Optimization in Unstructured Document Analysis

T0 review · 5 major / 5 minor · reviewed 2026-08-06 · deepseek-v4-flash

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

Pith's one-line read QUEST optimizes LLM-powered document analysis by retrieving only relevant segments, building a fresh execution plan per document, and converting joins into filters, reporting 30%–6× cost savings and 10–27% higher F1.

desk verdict Fresh empirical system, but the central theoretical guarantee (Lemma 2) is false under the paper's own cost model, and the proof is deferred to an anonymous tech report. read the letter →

arxiv 2507.06515 v2 pith:I4GABHR5 submitted 2025-07-09 cs.DB

Zhaoze Sun , Qiyan Deng , Chengliang Chai , Kaisen Jin , Xinyu Guo , Han Han , Ye Yuan , Guoren Wang

show 1 more author

Lei Cao

This is my paper

classification cs.DB

keywords queryoptimizationLLMcostreductionunstructureddocumentanalysisindex-basedextractionevidence-augmentedretrievalfilterorderingjointransformationinstance-optimizedexecution

verification ladder T0 review T1 audit T2 compute T3 formal

The pith

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

The reading

QUEST is a query optimizer for systems that analyze unstructured text documents with large language models. It argues that the cheapest way to run such a query is to retrieve only the text segments that can supply the requested attributes, then plan the remaining extraction per document rather than once for the whole query. Its execution-time optimizer orders filters by a priority score that balances selectivity with LLM token cost, and it turns joins into filters that can be reordered with the rest of the plan. The paper reports 30% to 6× cost savings and 10–27% higher F1 than state-of-the-art baselines across the LCR, WikiText, and SWDE datasets.

What carries the argument

The load-bearing mechanism is instance-optimized query execution. For a conjunction, each filter $\theta_k$ receives a priority $(1-p_k)/c_k$, where $p_k$ is its selectivity estimated from samples and $c_k$ is its LLM cost measured as the token count of retrieved segments; sorting by this score minimizes expected extraction cost, and the disjunction analogue uses $p_k/c_k$. The join transformation turns a join into an IN filter, and Lemma 2 guarantees that this plan costs no more than predicate pushdown. The two-level index (document-level, then segment-level) with evidence-augmented retrieval is what supplies the per-attribute cost $c_k$ by limiting which text the LLM sees.

What would settle it

Construct a corpus where a target attribute's value can only be derived by combining facts from two separated segments—for example, a birth date in one paragraph and the current date in a later one—and run a query for that attribute; if QUEST's recall collapses to near zero while a full-document extractor succeeds, the single-segment assumption is the breaking point.

Watch

Extended reading notes

Core claim

The paper's central claim is that the high cost of LLM-based extraction over unstructured documents is not inevitable: a query optimizer can cut it sharply without hurting accuracy. On the retrieval side, a two-level index first filters out irrelevant documents, then retrieves only the semantically coherent segments that plausibly contain an attribute; sampled evidence and LLM-synthesized evidence augment the query embedding so that value-bearing segments are not missed. On the execution side, QUEST abandons one-plan-per-query in favor of per-document plans: filters are sorted by a cost-aware priority, and a join is transformed into an IN filter that is ordered with the other filters. The paper states Lemma 2, that this join-to-filter plan is always at least as cheap as predicate pushdown, and reports experiments in which QUEST beats existing systems in both token cost and F1.

Load-bearing premise

The central assumption is that every attribute value sought by a query is fully contained in one semantically coherent text segment, so one retrieved segment suffices for extraction.

Editorial extensions

If this is right

Queries over thousands of documents become affordable because each document's plan invokes the LLM only until the filters decide the tuple's fate.
Join queries over text no longer need to extract both join attributes up front; the join can be applied as a filter after extracting only one side.
Users can express SQL-style queries over documents without hand-engineering prompts, because sampling collects the evidence needed to find attribute-bearing segments automatically.
The savings should hold across different LLMs because the optimizer reduces token counts and call frequency rather than relying on any one model's behavior.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

A natural stress test is to vary the 5% sampling ratio and the cluster-count $K$ on document types far outside the three datasets; the reported margins would narrow or widen depending on how stable retrieval evidence is across domains.
The single-segment assumption suggests an immediate extension: when an attribute is not found in any single retrieved segment, QUEST could retrieve neighboring segments and re-run extraction, with the added token cost priced into the optimizer.
If the approach's cost model is right, the same per-document planning should extend to approximate aggregation over samples, where the optimizer chooses how many documents to read based on each document's estimated extraction cost.

Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Desk Editor's Note

read the letter

Here's the short version: the paper's empirical system is plausible and the combination of ideas is new, but the central theoretical guarantee—Lemma 2, that join-to-filter always beats predicate pushdown—is false under the paper's own cost model, and the proof is not in the text.

What's genuinely new: QUEST produces per-document execution plans instead of one plan per query, and it augments segment retrieval with evidence collected from sampled documents. The two-level index (document + segment) is a sensible way to cut LLM tokens. The join-to-filter transformation is a reasonable heuristic, and the experiments show 30%-6x token savings and F1 gains over ZenDB, Lotus, Palimpzest, and a RAG baseline. Those results are suggestive.

The load-bearing problem is Lemma 2. Compare Plan 1 (filter first, then extract join attribute) with Plan 2 (extract join attribute first, then filter) on table T2. Using the paper's own equations (7) and (9), the difference is (1-p2)*sum(c_{a'}) - (1-p_{a'})*sum(C_2). With sum(c_{a'})=1000, sum(C_2)=1, p2=0.1, p_{a'}=1, Plan 2 costs 900 more than Plan 1. So the lemma simply is not true. The proof is deferred to an anonymous technical report, which prevents checking.

Other concerns, in decreasing order of importance. The filter-ordering evaluation compares QUEST to Exhaust using the same sample-estimated costs and selectivities, so matching Exhaust shows internal consistency, not real optimality. Distance thresholds are fit on the evaluation data. There are no error bars, no code release, no LLM versions or seeds. And Section 4.1 explicitly assumes each attribute can be extracted from a single segment; that's a real limitation for attributes that span multiple segments, and it likely bounds recall.

Minor: the priority rule (1-p)/c is Smith's rule from scheduling; a citation would be appropriate, but this is a missed connection rather than an error.

Bottom line: this is a paper for the LLM-powered data systems community. The empirical system and the per-document optimization idea deserve discussion. But the false lemma and missing artifacts mean the paper is not ready as stated. A serious editor should send it to review—there is enough substance here—but the review should require either a correct proof, a weakened claim, or a clear heuristic framing, plus code/data and uncertainty quantification.

Referee Report

5 major / 5 minor

Summary. The paper proposes QUEST, a system for SPJ queries over unstructured text documents in which attribute values are extracted by LLMs. It combines two lines of optimization: (i) an index-based extraction component with a document-level index, a segment-level index, and evidence-augmented retrieval, intended to reduce the token cost of each extraction; and (ii) an execution-time, per-document query optimizer that orders filters by estimated LLM cost and selectivity, transforms joins into IN filters, and chooses join orders progressively during execution. Experiments on three datasets (LCR, WikiText, SWDE) compare QUEST with ZenDB, Palimpzest, Lotus, RAG, ClosedIE, and Evaporate, reporting cost savings and F1 improvements, and include ablations of the index, evidence, threshold, sampling ratio, and cluster count.

Significance. The problem is timely: LLM inference is the dominant cost in several emerging unstructured-data analytics systems, and there is relatively little work on cost-aware query optimization for such systems. The paper's ideas—document-level filtering before segment-level retrieval, evidence-augmented retrieval, per-document plans, and treating joins as filters—are plausible and worth investigating. The paper has concrete strengths: a simple, testable cost model; explicit priority rules for filter ordering; and ablations for the main components. However, the central theoretical guarantee advertised in the introduction and contributions (Lemma 2, the claim that join transformation is guaranteed to beat predicate pushdown) is false as stated under the paper's own cost model, and the proofs of the main lemmas are deferred to an anonymous technical report. The empirical evaluation is also currently too thin to support the headline quantitative claims: 500 documents, no error bars, no code release, and a filter-ordering comparison that mainly validates internal consistency.

major comments (5)

[§3.2.1, Lemma 2 and Eq. (9)] Lemma 2 is false as stated under the paper's own cost model. Suppressing the sum over documents, the T2 portion of Plan ① is C2 + p2·c_{a'}, while the T2 portion of Plan ②, as computed in the example after Eq. (9), is c_{a'} + p_{a'}·C2 (the IN filter is applied before θ2). The difference is (1−p2)c_{a'} − (1−p_{a'})C2, which is positive when the join attribute is expensive to extract and the IN filter does not prune; for c_{a'}=1000, C2=1, p2=0.1, p_{a'}=1, Plan ② costs 900 more than Plan ①. The unconditional "always" in Lemma 2 is therefore wrong. The proof is deferred to an anonymous technical report, and the "Selecting a Plan" heuristic based on the first two terms of Eqs. (9)–(10) inherits the same problem because the neglected third term can dominate. This is load-bearing for contribution (2), which advertises a guaranteed improvement over predicate pushdown.
[§3.1.2, §3.1.3, §3.2.1] The proofs of the optimality claims (Lemma 1, the disjunction variant, and Lemma 2) are all referred to an anonymous technical report [1] rather than provided in the paper. In a journal submission, the main optimality results must be proven in the paper or in a public appendix. The text also does not state the independence assumption implicit in multiplying selectivities in Eqs. (2) and (4); without that assumption, the claimed "optimal" order minimizes an approximate cost, not the true expected cost, and the approximation error can be large for correlated attributes.
[§5.3, Figure 6] The filter-ordering experiment compares QUEST with Exhaust using the same sample-estimated costs and selectivities for both algorithms. Matching Exhaust therefore demonstrates that QUEST reproduces the optimizer's internal optimum; it does not validate the cost model against true execution costs or against optimal plans computed from post-hoc ground-truth selectivities. The claim that QUEST "produces optimal orders in most cases" should be scoped to "optimal under the estimated cost model," and the paper should add a comparison with plans chosen using true selectivities or at least a sensitivity analysis of the estimates.
[§5.1–§5.2, Tables 2–3] The empirical evaluation is based on 500 documents total and reports a single number per metric in Tables 2–3 and Figures 4–8, with no error bars, confidence intervals, or repeated runs. The headline claims (30%–6× cost savings, 10–27% F1 improvement) are therefore not accompanied by any measure of dispersion. No code release is mentioned. At a minimum, the authors should provide per-query distributions or bootstrap intervals, and state whether code will be released for reproducibility.
[§4.1, §4.2, §5.5] Two threats to the retrieval accuracy claim are not addressed. First, Section 4.1 states the design goal that "each attribute can be extracted from a single segment," but the paper does not evaluate how often this assumption holds; attributes whose evidence is scattered across multiple segments or across the whole document will be missed, which would lower recall. Second, the automatic threshold selection in Section 4.2 uses sampled documents from the same query, and the ablation in Figure 8-c varies τ around the value selected for the evaluation query; if the same documents are used for tuning and evaluation, the accuracy results are optimistically biased. A separate tuning/evaluation split should be described.

minor comments (5)

[§3.2.1] The sentence introducing Lemma 2 says "Plans ① and ② are at least as good as Plan ①"; it should read "Plans ② and ③ are at least as good as Plan ①."
[§5.2] The text says Eva spends "a few tokes on generating code"; this should be "tokens."
[Figure 7] The captions do not define the E1–E3 and F1–F3 groups; the text defines them as selectivity ranges and filter-count groups, so the captions should be self-contained.
[§4.1] Please specify the exact versions of E5Model and SemanticChunker used, and the embedding dimension, to aid reproducibility.
[§3.1.4] The O(|θ| log |θ|) complexity claim for Algorithm 1 is stated without derivation; either prove it or describe the data structures used for the priority queue and expression tree.

Circularity Check

2 steps flagged · score 5.0 of 10

Optimality experiments pit QUEST against the same sample-fitted cost model that generates its plans, and the join-transformation guarantee is deferred to the authors' own technical report while contradicting the paper's own cost equations.

fitted input called prediction [Section 3.1.2 (Eqs. 2-3) and Section 5.3, Figure 6, 'Comparison of Filter Ordering Strategies']
"QUEST then uses its selectivity p_k estimated on the sampled table, and the cost c_k to compute a priority score ... Exhaust: It exhaustively enumerates all possible orders and returns the optimal one per document ... these methods are ranked as follows by the LLM cost: QUEST≈ Exhaust < Average_cost < Selectivity < Random. This shows that QUEST produces optimal orders in most cases."

Both QUEST's priority score and the Exhaust baseline's 'optimal' per-document order are computed from the same sample-derived parameters: QUEST uses selectivity p_k and cost c_k estimated on the sampled documents, and Exhaust returns the optimal order by enumerating the same cost model (Eqs. 2-3). Since Lemma 1 asserts that sorting by (1-p_k)/c_k minimizes exactly that cost function, the agreement QUEST≈Exhaust is guaranteed by construction given the lemma. The experiment therefore verifies the self-consistency of the sorting rule with the fitted estimates and cannot test whether the cost model reflects real LLM extraction costs.
self citation load bearing [Section 3.2.1, 'Selecting a Plan' (Eqs. 7-10) and Section 1 (contribution 2)]
"First, we present a lemma showing that Plans ① and ② are at least as good as Plan ① in all cases. The proof can be found in our technical report [1]. ... Lemma 2. Given a query Q containing a join operation, the expected cost of Plan ① is always greater than or equal to that of Plan ② and Plan ③."

The paper advertises 'a join transformation technique that is guaranteed to outperform the classical predicate pushdown strategy' and uses Lemma 2 as that guarantee, but the proof is not in the paper: it is deferred to reference [1], the authors' own anonymous 'Full_version' technical report of QUEST, so the derivation chain for the load-bearing theoretical claim terminates in a self-citation. The claim is also contradicted by the paper's own cost model: subtracting Plan ② (Eq. 9) from Plan ① (Eq. 7) gives (1-p_{a'})Σ_i C^i_2 - (1-p_2)Σ_i c^i_{a'}, which is negative when the join attribute is costly to extract (large Σ_i c^i_{a'}) and the IN filter is non-selective (p_{a'}=1); then Plan ② is strictly more expensive than Plan ①, the opposite of Lemma 2.

full rationale

The headline empirical claims (30%-6x cost savings and 10-27% F1 improvement) are evaluated against external systems (ZenDB, Palimpzest, Lotus, RAG, ClosedIE, Evaporate) on three real datasets with human-verified ground truth; those comparisons are self-contained and are not circular. Two load-bearing sub-claims are partially circular. First, the filter-ordering evaluation (Sec. 5.3, Figure 6) compares QUEST with Exhaust using the same sample-estimated selectivities and costs that drive QUEST's own priority score; because Lemma 1 asserts that the sort rule minimizes that same empirical cost function, the match checks internal consistency rather than real optimality. Second, Lemma 2, the advertised guarantee that join-to-IN-filter plans always beat predicate pushdown, has its proof deferred to reference [1], the authors' own anonymous full version, and the lemma fails under the paper's own Eqs. 7 and 9 whenever the join attribute is expensive to extract and the IN filter is not selective; the theorem thus rests on an unverified and contradicted self-citation. A minor in-sample component exists in the threshold ablation (Sec. 4.2 and Fig. 8-c), where the reported F1 curve includes the sampled documents used to fit tau and gamma_i; the effect is limited because the fit uses only the approximately 5% sample. Lemma 1's proof is likewise deferred to [1] rather than derived in the paper, although the (1-p)/c ordering rule itself is a standard expensive-predicate result. Overall, the central derivation is not equivalent to its inputs, but the optimality validation loop and the join guarantee are partly self-referential, yielding score 5.

Assumptions & free parameters 5 free parameters · 5 assumptions · 0 invented entities

The central cost model relies on sample-derived parameters: sampling ratio, cluster count, synthetic evidence count, and two distance thresholds. No new physical entities are introduced. The main non-empirical inputs are representativeness and single-segment extraction assumptions.

free parameters (5)

sampling_ratio = 5%
Fraction of documents sampled for selectivity, cost, and threshold estimation (Sec 4.2); ablation in Fig 8d tunes around this default.
evidence_cluster_count_k = 3
Number of k-means clusters used to compress evidence embeddings (Sec 4.2); ablation in Fig 8e.
synthetic_evidence_count = 20
Number of LLM-synthesized text segments used when no sampled evidence is found for an attribute (Sec 4.2).
document_threshold_tau = auto-adjusted per query from sample max distance
Distance threshold for document retrieval is initialized high then set to the maximum embedding distance of relevant sampled documents (Sec 4.2).
segment_threshold_gamma_i = max pairwise distance among positive sampled segments plus 0.1
Distance threshold for segment retrieval per attribute, fit to sampled documents and then increased by 0.1 (Sec 4.2, Sec 5.5).

assumptions (5)

domain assumption Every attribute value appears in a single segment
Stated as a design goal in Sec 4.1; if false, segment-level extraction misses values.
domain assumption Embedding similarity indicates attribute relevance
Used for document and segment retrieval in Sec 4.2; no calibration against false positives is provided.
domain assumption The 5% sampled documents are representative for selectivity and threshold estimation
All cost model parameters and thresholds come from the sample (Sec 2.2, Sec 4.2).
ad hoc to paper Filter outcomes are independent in the cost model
Eq 2 multiplies selectivities; no correlation modeling is included, which matters when attributes co-occur.
ad hoc to paper The expected cost model correctly ranks execution orders
Optimality proofs for Lemmas 1 and 2 are deferred to the anonymous technical report [1] and are not machine checked; Lemma 2 appears false in general.

how reviews work

0 comments

Cite this review

Pith. "Pith review of QUEST: Query Optimization in Unstructured Document Analysis." pith.science (2026). https://pith.science/paper/I4GABHR5

@misc{pith2026250706515,
  author       = {Pith},
  title        = {Pith review of: QUEST: Query Optimization in Unstructured Document Analysis},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/I4GABHR5}},
  note         = {Machine review of arXiv:2507.06515}
}

read the original abstract

Most recently, researchers have started building large language models (LLMs) powered data systems that allow users to analyze unstructured text documents like working with a database because LLMs are very effective in extracting attributes from documents. In such systems, LLM-based extraction operations constitute the performance bottleneck of query execution due to the high monetary cost and slow LLM inference. Existing systems typically borrow the query optimization principles popular in relational databases to produce query execution plans, which unfortunately are ineffective in minimizing LLM cost. To fill this gap, we propose QUEST, which features a bunch of novel optimization strategies for unstructured document analysis. First, we introduce an index-based strategy to minimize the cost of each extraction operation. With this index, QUEST quickly retrieves the text segments relevant to the target attributes and only feeds them to LLMs. Furthermore, we design an evidence-augmented retrieval strategy to reduce the possibility of missing relevant segments. Moreover, we develop an instance-optimized query execution strategy: because the attribute extraction cost could vary significantly document by document, QUEST produces different plans for different documents. For each document, QUEST produces a plan to minimize the frequency of attribute extraction. The innovations include LLM cost-aware operator ordering strategies and an optimized join execution approach that transforms joins into filters. Extensive experiments on 3 real-world datasets demonstrate the superiority of QUEST, achieving 30%-6x cost savings while improving the F1 score by 10% -27% compared with state-of-the-art baselines.

Figures

Figures reproduced from arXiv: 2507.06515 by the authors.

**Figure 1.** QUEST Framework document to document. To quickly produce these plans, QUEST features novel optimization strategies, which, although lightweight, guarantee to produce optimal plans in many cases. The query execution engine uses index-based data extraction and other analytical operations such as join to execute query plans efficiently. QUEST’s key components include: a two-level index-based strategy to minimize the… view at source ↗

**Figure 2.** Example of the Expression Tree. Example 3. For the expression in [PITH_FULL_IMAGE:figures/full_fig_p006_2.png] view at source ↗

**Figure 3.** Example for JOIN 6 [PITH_FULL_IMAGE:figures/full_fig_p006_3.png] view at source ↗

Figures from the paper (5 more)

**Figure 4.** Figure 4: F1-Score of Baselines (Different Query Groups) [PITH_FULL_IMAGE:figures/full_fig_p010_4.png]

**Figure 5.** Figure 5: Cost of Baselines (Different Query Groups) [PITH_FULL_IMAGE:figures/full_fig_p010_5.png]

**Figure 7.** Figure 7: Evaluation of Join the joins. We also compare QUEST with the Optimal plan, which is obtained by assuming the selectivity of each filter is known. We create queries incorporating one of the three aforementioned joins, along with some random filters. In total, we constru…

**Figure 6.** Figure 6: Comparison of Filter Reordering Strategies [PITH_FULL_IMAGE:figures/full_fig_p011_6.png]

**Figure 8.** Figure 8: Ablation Studies In [PITH_FULL_IMAGE:figures/full_fig_p012_8.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

46 extracted references · 23 canonical work pages

[1]

https://anonymous.4open.science/r/QUEST/Full_version.pdf

[n.d.]. https://anonymous.4open.science/r/QUEST/Full_version.pdf

work page
[2]

https://solutionsreview.com/data-management/80-percent-of-your-data- will-be-unstructured-in-five-years/

2019. https://solutionsreview.com/data-management/80-percent-of-your-data- will-be-unstructured-in-five-years/

work page 2019
[3]

Simran Arora, Brandon Yang, Sabri Eyuboglu, Avanika Narayan, Andrew Hojel, Immanuel Trummer, and Christopher Ré. 2023. Language Models Enable Simple Systems for Generating Structured Views of Heterogeneous Data Lakes. Proc. VLDB Endow. 17, 2 (Oct. 2023), 92–105. https://doi.org/10.14778/3626292.3626294

arXiv 2023
[4]

Dhananjay Ashok and ZacharyC. Lipton. 2023. PromptNER: Prompting For Named Entity Recognition. (May 2023)

work page 2023
[5]

Parikshit Bansal and Amit Sharma. 2023. Large language models as annota- tors: Enhancing generalization of nlp models at minimal cost. arXiv preprint arXiv:2306.15766 (2023)

arXiv 2023
[6]

Deng Cai, Yan Wang, Lemao Liu, and Shuming Shi. 2022. Recent Advances in Retrieval-Augmented Text Generation. InProceedings of the 45th International ACM SIGIR Conference on Research and Development in Information Retrieval . https://doi.org/10.1145/3477495.3532682

work page arXiv 2022
[7]

Surajit Chaudhuri. 1998. An overview of query optimization in relational systems. In Proceedings of the seventeenth ACM SIGACT-SIGMOD-SIGART symposium on Principles of database systems . 34–43

work page 1998
[8]

Zui Chen, Zihui Gu, Lei Cao, Ju Fan, Sam Madden, and Nan Tang. [n.d.]. Sym- phony: Towards Natural Language Query Answering over Multi-modal Data Lakes. ([n. d.])

work page

Show all 46 references

[10]

Hanjun Dai, Bethany Yixin Wang, Xingchen Wan, Bo Dai, Sherry Yang, Azade Nova, Pengcheng Yin, Phitchaya Mangpo Phothilimthana, Charles Sutton, and Dale Schuurmans. 2024. UQE: A Query Engine for Unstructured Databases. In The Thirty-eighth Annual Conference on Neural Informatio...

2024
[11]

Zhicheng Ding, Jiahao Tian, Zhenkai Wang, Jinman Zhao, and Siyang Li. 2024. Data imputation using large language model to accelerate recommendation system. arXiv preprint arXiv:2407.10078 (2024)

2024 arXiv
[12]

Meihao Fan, Xiaoyue Han, Ju Fan, Chengliang Chai, Nan Tang, Guoliang Li, and Xiaoyong Du. 2024. Cost-effective in-context learning for entity resolution: A design space exploration. In 2024 IEEE 40th International Conference on Data Engineering (ICDE). IEEE, 3696–3709

2024
[13]

Wenqi Fan, Yujuan Ding, Liangbo Ning, Shijie Wang, Hengyun Li, Dawei Yin, Tat-Seng Chua, and Qing Li. 2024. A survey on rag meeting llms: Towards retrieval-augmented large language models. In Proceedings of the 30th ACM SIGKDD Conference on Knowledge Discovery and Data Mining ...

2024
[14]

Filippo Galgani and Achim Hoffmann. 2010. LEXA: Towards Automatic Legal Citation Classification. InAI 2010: Advances in Artificial Intelligence (Lecture Notes in Computer Science), Jiuyong Li (Ed.), Vol. 6464. Springer Berlin Heidelberg, 445 –454

2010
[15]

Dawei Gao, Haibin Wang, Yaliang Li, Xiuyu Sun, Yichen Qian, Bolin Ding, and Jingren Zhou. [n.d.]. Text-to-SQL Empowered by Large Language Models: A Benchmark Evaluation. ([n. d.])
[16]

Yingqi Gao, Yifu Liu, Xiaoxia Li, Xiaorong Shi, Yin Zhu, Yiming Wang, Shiqi Li, Wei Li, Yuntao Hong, Zhiling Luo, et al. 2024. XiYan-SQL: A Multi-Generator Ensemble Framework for Text-to-SQL. arXiv preprint arXiv:2411.08599 (2024)

2024 arXiv
[17]

Yunfan Gao, Yun Xiong, Xinyu Gao, Kangxiang Jia, Jinliu Pan, Yuxi Bi, Yi Dai, Jiawei Sun, and Haofen Wang. 2023. Retrieval-augmented generation for large language models: A survey. arXiv preprint arXiv:2312.10997 (2023)

2023 arXiv
[18]

Tiezheng Ge, Kaiming He, Qifa Ke, and Jian Sun. 2013. Optimized product quantization. IEEE transactions on pattern analysis and machine intelligence 36, 4 (2013), 744–755

2013
[19]

Qiang Hao, Rui Cai, Yanwei Pang, and Lei Zhang. 2011. From one tree to a forest. In Proceedings of the 34th international ACM SIGIR conference on Research and development in Information Retrieval . https://doi.org/10.1145/2009916.2010020

2011
[20]

Ahatsham Hayat and Mohammad Rashedul Hasan. 2024. CLAIM Your Data: Enhancing Imputation Accuracy with Contextual Large Language Models. arXiv preprint arXiv:2405.17712 (2024)

2024 arXiv
[21]

Pengcheng He, Jianfeng Gao, and Weizhu Chen. 2021. Debertav3: Improving deberta using electra-style pre-training with gradient-disentangled embedding sharing. arXiv preprint arXiv:2111.09543 (2021)

2021 arXiv
[22]

Pengcheng He, Xiaodong Liu, Jianfeng Gao, and Weizhu Chen. 2020. DeBERTa: Decoding-enhanced BERT with Disentangled Attention. Cornell University - arXiv,Cornell University - arXiv (Jun 2020)

2020
[23]

Pengcheng He, Xiaodong Liu, Jianfeng Gao, and Weizhu Chen. 2021. DEBERTA: DECODING-ENHANCED BERT WITH DISENTANGLED ATTENTION. In Inter- national Conference on Learning Representations . https://openreview.net/forum? id=XPZIaotutsD

2021
[24]

Yizhu Jiao, Ming Zhong, Sha Li, Ruining Zhao, Siru Ouyang, Heng Ji, and Jiawei Han. 2023. Instruct and extract: Instruction tuning for on-demand information extraction. arXiv preprint arXiv:2310.16040 (2023)

2023 arXiv
[25]

Saehan Jo and Immanuel Trummer. 2024. ThalamusDB: Approximate Query Processing on Multi-Modal Data. Proceedings of the ACM on Management of Data 2, 3 (2024), 1–26

2024
[26]

Patrick Lewis, Ethan Perez, Aleksandra Piktus, Fabio Petroni, Vladimir Karpukhin, Naman Goyal, Heinrich Küttler, Mike Lewis, Wen-tau Yih, Tim Rocktäschel, Sebastian Riedel, and Douwe Kiela. 2020. Retrieval-augmented generation for knowledge-intensive NLP tasks. In Proceedings ...

2020
[27]

Huayang Li, Yixuan Su, Deng Cai, Yan Wang, and Lemao Liu. [n.d.]. A Survey on Retrieval-Augmented Text Generation. ([n. d.])
[28]

Peng Li, Yeye He, Dror Yashar, Weiwei Cui, Song Ge, Haidong Zhang, Danielle Rifinski Fainman, Dongmei Zhang, and Surajit Chaudhuri. 2023. Table- gpt: Table-tuned gpt for diverse table tasks. arXiv preprint arXiv:2310.09263 (2023)

2023 arXiv
[29]

Yiming Lin, Madelon Hulsebos, Ruiying Ma, Shreya Shankar, Sepanta Zeigham, Aditya G Parameswaran, and Eugene Wu. 2024. Towards Accurate and Ef- ficient Document Analytics with Large Language Models. arXiv preprint arXiv:2405.04674 (2024)

2024 arXiv
[30]

Chunwei Liu, Matthew Russo, Michael Cafarella, Lei Cao, Peter Baile Chen, Zui Chen, Michael Franklin, Tim Kraska, Samuel Madden, Rana Shahout, and Gerardo Vitagliano. [n.d.]. Palimpzest: Optimizing AI-Powered Analytics with Declarative Query Processing. In Proceedings of the C...

2025
[31]

Xinyu Liu, Shuyu Shen, Boyan Li, Peixian Ma, Runzhi Jiang, Yuxin Zhang, Ju Fan, Guoliang Li, Nan Tang, and Yuyu Luo. 2024. A Survey of NL2SQL with Large Language Models: Where are we, and where are we going? arXiv preprint arXiv:2408.05109 (2024)

2024
[32]

Yurong Liu, Eduardo Pena, Aecio Santos, Eden Wu, and Juliana Freire. 2024. Magneto: Combining Small and Large Language Models for Schema Matching. arXiv:2412.08194 [cs.DB] https://arxiv.org/abs/2412.08194

2024 arXiv
[33]

Yu A Malkov and Dmitry A Yashunin. 2018. Efficient and robust approximate nearest neighbor search using hierarchical navigable small world graphs. IEEE transactions on pattern analysis and machine intelligence 42, 4 (2018), 824–836

2018
[34]

Peeters, and Stijn Vansummeren

Marcel Parciak, Brecht Vandevoort, Frank Neven, Liesbet M. Peeters, and Stijn Vansummeren. 2024. Schema Matching with Large Language Models: an Experi- mental Study. arXiv:2407.11852 [cs.DB] https://arxiv.org/abs/2407.11852

2024 arXiv
[35]

Liana Patel, Siddharth Jha, Carlos Guestrin, and Matei Zaharia. 2024. LOTUS: En- abling Semantic Queries with LLMs Over Tables of Unstructured and Structured Data. arXiv preprint arXiv:2407.11418 (2024)

2024 arXiv
[36]

PengLi and TianxiangSun ect all. 2023. CodeIE: Large Code Generation Models are Better Few-Shot Information Extractors. (May 2023)

2023
[37]

Michał Pietruszka, Michał Turski, Łukasz Borchmann, Tomasz Dwojak, Gabriela Pałka, Karolina Szyndler, Dawid Jurkiewicz, and Łukasz Garncarek. 2022. Sta- ble: Table generation framework for encoder-decoder models. arXiv preprint arXiv:2206.04045 (2022)

2022 arXiv
[38]

Bruno R Preiss. 1999. Data structures and algorithms . John Wiley & Sons, Inc

1999
[39]

Oscar Sainz, Iker Garcia-Ferrero, Rodrigo Agerri, OierLopezde Lacalle, German Rigau, and Eneko Agirre. 2023. GoLLIE: Annotation Guidelines improve Zero- Shot Information-Extraction. (Oct 2023)

2023
[40]

Parth Sarthi, Salman Abdullah, Aditi Tuli, Shubh Khanna, Anna Goldie, and ChristopherD. Manning. 2024. RAPTOR: Recursive Abstractive Processing for Tree-Organized Retrieval. (Jan 2024)

2024
[41]

James Thorne, Majid Yazdani, Marzieh Saeidi, Fabrizio Silvestri, Sebastian Riedel, and Alon Halevy. 2021. From natural language processing to neural databases. Proceedings of the VLDB Endowment (Feb 2021), 1033–1039. https://doi.org/10. 14778/3447689.3447706

2021
[42]

Matthias Urban and Carsten Binnig. [n.d.]. Towards Multi-Modal DBMSs for Seamless Querying of Texts and Tables. ([n. d.])
[43]

Matthias Urban and Carsten Binnig. 2023. CAESURA: Language Models as Multi-Modal Query Planners. arXiv preprint arXiv:2308.03424 (2023)

2023 arXiv
[44]

Liang Wang, Nan Yang, Xiaolong Huang, Jiao Binxing, Linjun Yang, Daxin Jiang, Rangan Majumder, and Furu Wei. 2022. Text Embeddings by Weakly-Supervised Contrastive Pre-training. Cornell University - arXiv,Cornell University - arXiv (Dec 2022)

2022
[45]

Xueqing Wu, Jiacheng Zhang, and Hang Li. 2022. Text-to-Table: A New Way of Information Extraction. In Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers) . https://doi. org/10.18653/v1/2022.acl-long.180

2022 doi
[46]

Ruixuan Xiao, Yiwen Dong, Junbo Zhao, Runze Wu, Minmin Lin, Gang Chen, and Haobo Wang. 2023. FreeAL: Towards Human-Free Active Learning in the Era of Large Language Models. In Proceedings of the 2023 Conference on Empirical Methods in Natural Language Processing . 14520–14535

2023
[47]

Haochen Zhang, Yuyang Dong, Chuan Xiao, and Masafumi Oyamada. 2024. Jellyfish: Instruction-Tuning Local Large Language Models for Data Preprocess- ing. In Proceedings of the 2024 Conference on Empirical Methods in Natural Lan- guage Processing, Yaser Al-Onaizan, Mohit Bansal, ...

2024 doi

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.

[1] [1]

https://anonymous.4open.science/r/QUEST/Full_version.pdf

[n.d.]. https://anonymous.4open.science/r/QUEST/Full_version.pdf

work page

[2] [2]

https://solutionsreview.com/data-management/80-percent-of-your-data- will-be-unstructured-in-five-years/

2019. https://solutionsreview.com/data-management/80-percent-of-your-data- will-be-unstructured-in-five-years/

work page 2019

[3] [3]

Simran Arora, Brandon Yang, Sabri Eyuboglu, Avanika Narayan, Andrew Hojel, Immanuel Trummer, and Christopher Ré. 2023. Language Models Enable Simple Systems for Generating Structured Views of Heterogeneous Data Lakes. Proc. VLDB Endow. 17, 2 (Oct. 2023), 92–105. https://doi.org/10.14778/3626292.3626294

arXiv 2023

[4] [4]

Dhananjay Ashok and ZacharyC. Lipton. 2023. PromptNER: Prompting For Named Entity Recognition. (May 2023)

work page 2023

[5] [5]

Parikshit Bansal and Amit Sharma. 2023. Large language models as annota- tors: Enhancing generalization of nlp models at minimal cost. arXiv preprint arXiv:2306.15766 (2023)

arXiv 2023

[6] [6]

Deng Cai, Yan Wang, Lemao Liu, and Shuming Shi. 2022. Recent Advances in Retrieval-Augmented Text Generation. InProceedings of the 45th International ACM SIGIR Conference on Research and Development in Information Retrieval . https://doi.org/10.1145/3477495.3532682

work page arXiv 2022

[7] [7]

Surajit Chaudhuri. 1998. An overview of query optimization in relational systems. In Proceedings of the seventeenth ACM SIGACT-SIGMOD-SIGART symposium on Principles of database systems . 34–43

work page 1998

[8] [8]

Zui Chen, Zihui Gu, Lei Cao, Ju Fan, Sam Madden, and Nan Tang. [n.d.]. Sym- phony: Towards Natural Language Query Answering over Multi-modal Data Lakes. ([n. d.])

work page