pith. sign in

arxiv: 2606.31808 · v1 · pith:UDTYJBMCnew · submitted 2026-06-30 · 💻 cs.AI · cs.DB

Large Databases Need Small, Open-Weight Language Models

Parker Glenn , Alfy Samuel This is my paper

Pith reviewed 2026-07-01 05:26 UTC · model grok-4.3

classification 💻 cs.AI cs.DB
keywords open-weight language modelsquantized modelsLM-database integrationrelational operatorscost reductionlocal deploymentBlendSQL
0
0 comments X

The pith

Quantized open-weight models on 16GB VRAM match closed-source accuracy for database tasks at lower cost and latency.

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

The paper shows that proprietary API costs for language-model-enhanced database operations on large datasets quickly become unsustainable, often surpassing $10,000 for a single experimental run. It establishes that quantized open-weight models running locally can reach or surpass the accuracy of closed-source models while cutting both latency and expense. Through optimizations and integration with the BlendSQL v0.1.0 framework, the work records a 390x cost reduction and 3.8x latency improvement, indicating that closed-source APIs are not required for effective LM-database integration.

Core claim

Quantized, open-weight models running locally on just 16GB of VRAM can match or exceed the accuracy of closed-source counterparts at lower latency and a fraction of the price for LM-enhanced relational operators on large databases, demonstrated by their deployment in the BlendSQL v0.1.0 framework that yields a 390x reduction in overall costs and 3.8x reduction in latency.

What carries the argument

System optimizations required to efficiently deploy open-weight models locally within an LM-DB system, integrated into the BlendSQL framework.

If this is right

  • LM-enhanced relational operators on large databases become feasible without incurring prohibitive API expenses.
  • Thorough experimentation and practical deployment of LM-database systems can occur at a small fraction of prior costs.
  • Local model execution supplies lower latency than remote API calls for the same relational operators.
  • Open-weight models can serve as direct replacements for proprietary APIs in this application class.

Where Pith is reading between the lines

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

  • The same local-model strategy may transfer to other high-volume token-consumption settings outside databases.
  • Additional hardware-constrained environments could benefit from similar quantization and deployment tuning.
  • Testing the approach on a wider variety of database schemas and query patterns would test its generality.

Load-bearing premise

The specific tasks, datasets, and accuracy metrics used inside the BlendSQL v0.1.0 framework are representative of the broader class of LM-enhanced relational operators on large databases.

What would settle it

A set of experiments on different LM-enhanced database tasks or datasets where the local quantized models produce lower accuracy than the closed-source models would disprove the central claim.

Figures

Figures reproduced from arXiv: 2606.31808 by Alfy Samuel, Parker Glenn.

Figure 1
Figure 1. Figure 1: Plotting the quality, latency, and cost over the five [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Benchmarking the quality and latency of various LM-DB systems with open-weight LMs on the text-only [PITH_FULL_IMAGE:figures/full_fig_p005_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Plotting the average, minimum, and maximum CPU [PITH_FULL_IMAGE:figures/full_fig_p006_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Measuring the relative quality and latency change when incrementally adding query processing features to [PITH_FULL_IMAGE:figures/full_fig_p008_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Evaluating the impact of concurrency limits with LM inference requests on a single 16GB RTX 5080. This controls the [PITH_FULL_IMAGE:figures/full_fig_p008_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: While a properly harnessed quantized Gemma 4 [PITH_FULL_IMAGE:figures/full_fig_p009_6.png] view at source ↗
Figure 7
Figure 7. Figure 7: Full SemBench results. BlendSQL uses gemma-4-E4B-FP8, other systems use gemini-2.5-flash. (a) Full mmqa results. BlendSQL BigQuery LOTUS Palimpzest ThalamusDB Query Lat. (s) Quality Cost (5 runs) Lat. (s) Quality Cost (5 runs) Lat. (s) Quality Cost (5 runs) Lat. (s) Quality Cost (5 runs) Lat. (s) Quality Cost (5 runs) Q1 0.07 1.00 $2e-5 14.10 1.00 $0.05 7.40 1.00 $0.10 4.50 1.00 $0.01 – – – Q2a 55.15 1.00 … view at source ↗
Figure 9
Figure 9. Figure 9: Plotting the query-level latency and quality changes on SemBench scenarios with [PITH_FULL_IMAGE:figures/full_fig_p015_9.png] view at source ↗
read the original abstract

Language model systems built around proprietary APIs often operate on a token-based cost model. This becomes prohibitively expensive in the context of large databases, where LM-enhanced relational operators can incur costs exceeding $10,000 for a single set of experiments, hindering thorough research and practical deployment. In this paper, we demonstrate that quantized, open-weight models running locally on just 16GB of VRAM can match or exceed the accuracy of closed-source counterparts at lower latency and a fraction of the price, challenging the prevailing assumption that closed-source LM APIs are necessary for effective LM-database integration. We present and analyze the key system optimizations required to efficiently deploy these open-weight models within an LM-DB system. By integrating these local models into the BlendSQL v0.1.0 framework, we demonstrate a 390x reduction in overall costs and 3.8x reduction in latency compared to a proprietary LM API. We make our code available at https://github.com/CapitalOne-Research/play-by-the-type-rules/tree/main/sembench.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

2 major / 1 minor

Summary. The paper claims that quantized open-weight language models running locally on 16GB VRAM can match or exceed closed-source LM API accuracy for LM-enhanced relational operators on large databases, at lower latency and far lower cost. It reports a 390x overall cost reduction and 3.8x latency reduction when integrating these models into the BlendSQL v0.1.0 framework, along with the system optimizations needed for efficient local deployment, and releases the code.

Significance. If the accuracy equivalence is shown to hold under representative conditions, the result would meaningfully lower barriers to LM-DB research and deployment by removing dependence on token-priced proprietary APIs. The open release of code is a concrete strength for reproducibility.

major comments (2)
  1. [Abstract] Abstract: the numerical claims (accuracy match/exceed, 390x cost reduction, 3.8x latency) are presented without any description of the tasks, datasets, controls, error bars, or statistical tests used to obtain them, so the central performance assertions cannot be evaluated from the given text.
  2. [BlendSQL v0.1.0 framework] BlendSQL v0.1.0 integration section: the accuracy equivalence between quantized open-weight models and closed-source APIs is demonstrated only inside this framework; the manuscript supplies no argument or additional experiments showing that the chosen predicates, table sizes, or accuracy definitions are representative of the general class of LM-enhanced relational operators on large databases.
minor comments (1)
  1. [Abstract] The GitHub link in the abstract ends in a tree path; ensure the final URL resolves to the exact code and data used for the reported BlendSQL experiments.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for the constructive feedback. We address each major comment below and indicate planned revisions to the manuscript.

read point-by-point responses

  1. Referee: [Abstract] Abstract: the numerical claims (accuracy match/exceed, 390x cost reduction, 3.8x latency) are presented without any description of the tasks, datasets, controls, error bars, or statistical tests used to obtain them, so the central performance assertions cannot be evaluated from the given text.

    Authors: We agree that the abstract would be improved by additional context. In the revised manuscript we will expand the abstract to briefly describe the BlendSQL tasks (LM-enhanced relational operators such as semantic filters and joins), the datasets and table sizes used, and note that accuracy, cost, and latency results include error bars from repeated runs with full experimental controls and details reported in the main body. revision: yes

  2. Referee: [BlendSQL v0.1.0 framework] BlendSQL v0.1.0 integration section: the accuracy equivalence between quantized open-weight models and closed-source APIs is demonstrated only inside this framework; the manuscript supplies no argument or additional experiments showing that the chosen predicates, table sizes, or accuracy definitions are representative of the general class of LM-enhanced relational operators on large databases.

    Authors: BlendSQL is explicitly constructed to realize LM-enhanced relational operators, and the predicates, table sizes, and accuracy metrics in our experiments follow the framework's standard usage patterns for large-database scenarios. We will add an explicit discussion paragraph arguing for representativeness based on the framework design and typical LM-DB workloads. We do not provide experiments in other frameworks, as that would be outside the paper's scope; the contribution centers on showing local-model viability within an established LM-DB system. revision: partial

Circularity Check

0 steps flagged

No circularity: purely empirical measurements with no derivation chain

full rationale

The paper presents direct empirical comparisons of quantized open-weight models versus closed-source APIs inside the BlendSQL v0.1.0 framework, reporting measured accuracy equivalence, 390x cost reduction, and 3.8x latency reduction. No equations, fitted parameters, predictions, or uniqueness theorems are invoked; the central claims rest on experimental runs rather than any reduction of outputs to inputs by construction. Self-citations are absent from the provided text, and the work is self-contained against external benchmarks.

Axiom & Free-Parameter Ledger

0 free parameters · 0 axioms · 0 invented entities

The central claim rests on empirical measurements of accuracy, latency, and monetary cost for a fixed set of models and tasks inside BlendSQL. No free parameters, axioms, or invented entities are visible from the abstract.

pith-pipeline@v0.9.1-grok · 5700 in / 1069 out tokens · 14247 ms · 2026-07-01T05:26:51.226719+00:00 · methodology

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

66 extracted references · 29 canonical work pages · 10 internal anchors

  1. [1]

    Florian Angermeir, Maximilian Amougou, Mark Kreitz, Andreas Bauer, Matthias Linhuber, Davide Fucci, Daniel Mendez, Tony Gorschek, et al. 2025. Reflections on the Reproducibility of Commercial LLM Performance in Empirical Software Engineering Studies.arXiv preprint arXiv:2510.25506(2025)

    work page arXiv 2025

  2. [2]

    Samuel Arch, Yuchen Liu, Todd C Mowry, Jignesh M Patel, and Andrew Pavlo

  3. [3]

    The key to effective udf optimization: Before inlining, first perform outlin- ing.Proceedings of the VLDB Endowment18, 1 (2024), 1–13

    2024

  4. [4]

    Seongsu Bae, Daeun Kyung, Jaehee Ryu, Eunbyeol Cho, Gyubok Lee, Sunjun Kweon, Jungwoo Oh, Lei Ji, Eric Chang, Tackeun Kim, et al . 2023. Ehrxqa: A multi-modal question answering dataset for electronic health records with chest x-ray images.Advances in Neural Information Processing Systems36 (2023), 3867–3880

    2023

  5. [5]

    Lingjiao Chen, Matei Zaharia, and James Zou. 2024. How is ChatGPT’s behavior changing over time?Harvard Data Science Review6, 2 (2024)

    2024

  6. [6]

    Zhoujun Cheng, Tianbao Xie, Peng Shi, Chengzu Li, Rahul Nadkarni, Yushi Hu, Caiming Xiong, Dragomir Radev, Mari Ostendorf, Luke Zettlemoyer, et al. 2023. Binding Language Models in Symbolic Languages. InInternational Conference on Learning Representations (ICLR 2023)(01/05/2023-05/05/2023, Kigali, Rwanda)

    2023

  7. [7]

    Electric Choice. 2026. Electricity Rates by State. https://www.electricchoice.com/ electricity-prices-by-state/. Accessed: 2026-05-28

    2026

  8. [8]

    Suparno Roy Chowdhury, Manan Roy Choudhury, Tejas Anvekar, Muham- mad Ali Khan, Kaneez Zahra Rubab Khakwani, Mohamad Bassam Sonbol, Ir- baz Bin Riaz, and Vivek Gupta. 2026. Diagnosis, Bad Planning & Reasoning. Treatment, SCOPE–Planning for Hybrid Querying over Clinical Trial Data.arXiv preprint arXiv:2604.25120(2026)

    work page internal anchor Pith review Pith/arXiv arXiv 2026

  9. [9]

    Daniel Deutsch, Shyam Upadhyay, and Dan Roth. 2019. A general-purpose algo- rithm for constrained sequential inference. InProceedings of the 23rd Conference on Computational Natural Language Learning (CoNLL). 482–492

    2019

  10. [10]

    Yixin Dong, Charlie F Ruan, Yaxing Cai, Ruihang Lai, Ziyi Xu, Yilong Zhao, and Tianqi Chen. 2024. Xgrammar: Flexible and efficient structured generation engine for large language models.arXiv preprint arXiv:2411.15100(2024)

    work page arXiv 2024

  11. [11]

    Anas Dorbani, Sunny Yasser, Jimmy Lin, and Amine Mhedhbi. 2025. Beyond Quacking: Deep Integration of Language Models and RAG into DuckDB. (2025)

    2025

  12. [12]

    Yannis Foufoulas, Theoni Palaiologou, and Alkis Simitsis. 2025. The UDFBench Benchmark for General-purpose UDF Queries.Proceedings of the VLDB Endow- ment18, 9 (2025), 2804–2817

    2025

  13. [13]

    Yannis Foufoulas and Alkis Simitsis. 2023. Efficient execution of user-defined functions in SQL queries.Proceedings of the VLDB Endowment16, 12 (2023), 3874–3877

    2023

  14. [14]

    Saibo Geng, Hudson Cooper, Michał Moskal, Samuel Jenkins, Julian Berman, Nathan Ranchin, Robert West, Eric Horvitz, and Harsha Nori. 2025. Json- schemabench: A rigorous benchmark of structured outputs for language models. arXiv preprint arXiv:2501.10868(2025)

    work page arXiv 2025

  15. [15]

    Saibo Geng, Martin Josifoski, Maxime Peyrard, and Robert West. 2023. Grammar- constrained decoding for structured NLP tasks without finetuning.arXiv preprint arXiv:2305.13971(2023)

    work page arXiv 2023

  16. [16]

    Parker Glenn, Parag Dakle, Liang Wang, and Preethi Raghavan. 2024. BlendSQL: A Scalable Dialect for Unifying Hybrid Question Answering in Relational Algebra. InFindings of the Association for Computational Linguistics ACL 2024. 453–466

    2024

  17. [17]

    Parker Glenn, Alfy Samuel, and Daben Liu. [n.d.]. Play by the Type Rules: Inferring Constraints for Small Language Models in Declarative Programs. In EurIPS 2025 Workshop: AI for Tabular Data

    2025

  18. [18]

    Google. [n.d.]. Announcing BigQuery-managed AI functions for better SQL. https://cloud.google.com/blog/products/data-analytics/sql-reimagined- for-the-ai-era-with-bigquery-ai-functions. Accessed: 2026-05-07

    2026

  19. [19]

    Google. 2018. What a week! 105 announcements from Google Cloud Next ’18. https://blog.google/innovation-and-ai/infrastructure-and-cloud/google- cloud/100-plus-announcements-from-google-cloud-next-18/. Accessed: 2026- 05-13

    2018

  20. [20]

    Guidance. 2023. Guidance: A language model programming framework. https: //github.com/guidance-ai/guidance. Accessed: 2025-08-11

    2023

  21. [21]

    Jiawei Han, Yongjian Fu, Wei Wang, Krzysztof Koperski, Osmar Zaiane, et al

  22. [22]

    DMQL: A data mining query language for relational databases. InProc. 1996 SiGMOD, Vol. 96. 27–34

    1996

  23. [23]

    Joe Hellerstein, Christopher Ré, Florian Schoppmann, Daisy Zhe Wang, Eugene Fratkin, Aleksander Gorajek, Kee Siong Ng, Caleb Welton, Xixuan Feng, Kun Li, et al. 2012. The MADlib analytics library or MAD skills, the SQL.arXiv preprint arXiv:1208.4165(2012). Large Databases Need Small, Open-Weight Language Models

    work page internal anchor Pith review Pith/arXiv arXiv 2012

  24. [24]

    Tomasz Imieliński and Aashu Virmani. 1999. MSQL: A query language for database mining.Data Mining and Knowledge Discovery3, 4 (1999), 373–408

    1999

  25. [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. [26]

    Rohit Khoja, Devanshu Gupta, Yanjie Fu, Dan Roth, and Vivek Gupta. 2025. Weaver: Interweaving SQL and LLM for Table Reasoning.arXiv preprint arXiv:2505.18961(2025)

    work page arXiv 2025

  27. [27]

    Kyoungmin Kim, Kijae Hong, Caglar Gulcehre, and Anastasia Ailamaki. 2024. Optimizing LLM Inference for Database Systems: Cost-Aware Scheduling for Concurrent Requests.arXiv preprint arXiv:2411.07447(2024)

    work page arXiv 2024

  28. [28]

    Gonzalez, Hao Zhang, and Ion Stoica

    Woosuk Kwon, Zhuohan Li, Siyuan Zhuang, Ying Sheng, Lianmin Zheng, Cody Hao Yu, Joseph E. Gonzalez, Hao Zhang, and Ion Stoica. 2023. Efficient Memory Management for Large Language Model Serving with PagedAtten- tion. InProceedings of the ACM SIGOPS 29th Symposium on Operating Systems Principles

    2023

  29. [29]

    Loïc Lannelongue, Jason Grealey, and Michael Inouye. 2021. Green algorithms: quantifying the carbon footprint of computation.Advanced science8, 12 (2021), 2100707

    2021

  30. [30]

    Jiale Lao, Andreas Zimmerer, Olga Ovcharenko, Tianji Cong, Matthew Russo, Gerardo Vitagliano, Michael Cochez, Fatma Özcan, Gautam Gupta, Thibaud Hottelier, et al. 2025. SemBench: A Benchmark for Semantic Query Processing Engines.arXiv preprint arXiv:2511.01716(2025)

    work page arXiv 2025

  31. [31]

    Yoonho Lee, Roshen Nair, Qizheng Zhang, Kangwook Lee, Omar Khattab, and Chelsea Finn. 2026. Meta-Harness: End-to-End Optimization of Model Harnesses. arXiv preprint arXiv:2603.28052(2026)

    work page internal anchor Pith review Pith/arXiv arXiv 2026

  32. [32]

    Xiaoxi Li, Wenxiang Jiao, Jiarui Jin, Guanting Dong, Jiajie Jin, Yinuo Wang, Hao Wang, Yutao Zhu, Ji-Rong Wen, Yuan Lu, et al . 2025. Deepagent: A general reasoning agent with scalable toolsets.arXiv preprint arXiv:2510.21618(2025)

    work page arXiv 2025

  33. [33]

    Yin Lin, Tianjing Zeng, Zhongjun Ding, Rong Zhu, Bolin Ding, HV Jagadish, and Jingren Zhou. 2026. SEMA-SQL: Beyond Traditional Relational Querying with Large Language Models.arXiv preprint arXiv:2604.23477(2026)

    work page internal anchor Pith review Pith/arXiv arXiv 2026

  34. [34]

    Benjamin Lipkin, Benjamin LeBrun, Jacob Hoover Vigly, João Loula, David R MacIver, Li Du, Jason Eisner, Ryan Cotterell, Vikash Mansinghka, Timothy J O’Donnell, et al. 2025. Fast Controlled Generation from Language Models with Adaptive Weighted Rejection Sampling.arXiv preprint arXiv:2504.05410(2025)

    work page arXiv 2025

  35. [35]

    Chunwei Liu, Matthew Russo, Michael Cafarella, Lei Cao, Peter Baile Chen, Zui Chen, Michael Franklin, Tim Kraska, Samuel Madden, Rana Shahout, et al. 2025. Palimpzest: Optimizing ai-powered analytics with declarative query processing. InProceedings of the Conference on Innovative Database Research (CIDR). 2

    2025

  36. [36]

    Shu Liu, Asim Biswal, Audrey Cheng, Xiangxi Mo, Shiyi Cao, Joseph E Gonza- lez, Ion Stoica, and Matei Zaharia. 2024. Optimizing llm queries in relational workloads.CoRR(2024)

    2024

  37. [37]

    Qiuyang Mang, Yufan Xiang, Hangrui Zhou, Runyuan He, Jiaxiang Yu, Hanchen Li, Aditya Parameswaran, and Alvin Cheung. 2026. Horrila: Cost-Based Placement of Semantic Operators in Hybrid Query Plans.arXiv preprint arXiv:2604.09944(2026)

    work page internal anchor Pith review Pith/arXiv arXiv 2026

  38. [38]

    Toby Mao. [n.d.]. SQLGlot: Python SQL Parser and Transpiler. https://github. com/tobymao/sqlglot

  39. [39]

    Rosa Meo, Giuseppe Psaila, Stefano Ceri, et al. 1996. A new SQL-like operator for mining association rules. InVLDB, Vol. 96. 122–133

    1996

  40. [40]

    Microsoft. 2017. What’s new in SQL Server 2017. https://learn.microsoft.com/en- us/sql/sql-server/what-s-new-in-sql-server-2017?view=sql-server-ver17. Ac- cessed: 2026-05-13

    2017

  41. [41]

    Niels Mündler, Jingxuan He, Hao Wang, Koushik Sen, Dawn Song, and Mar- tin Vechev. 2025. Type-Constrained Code Generation with Language Models. Proceedings of the ACM on Programming Languages9, PLDI (2025), 601–626

    2025

  42. [42]

    doi:10.5281/zenodo.3509134 , version =

    The pandas development team. 2020.pandas-dev/pandas: Pandas. https://doi. org/10.5281/zenodo.3509134

    work page doi:10.5281/zenodo.3509134 2020

  43. [43]

    Kanghee Park, Timothy Zhou, and Loris D’Antoni. 2025. Flexible and efficient grammar-constrained decoding.arXiv preprint arXiv:2502.05111(2025)

    work page arXiv 2025

  44. [45]

    Liana Patel, Siddharth Jha, Melissa Pan, Harshit Gupta, Parth Asawa, Carlos Guestrin, and Matei Zaharia. 2024. Semantic Operators: A Declarative Model for Rich, AI-based Data Processing.arXiv preprint arXiv:2407.11418(2024)

    work page arXiv 2024

  45. [46]

    Polars. 2022. Polars: Extremely fast Query Engine for DataFrames, written in Rust. https://github.com/pola-rs/polars. Accessed: 2026-04-23

    2022

  46. [47]

    2025.PostgreSQL: The World’s Most Advanced Open Source Relational Database

    PostgreSQL. 2025.PostgreSQL: The World’s Most Advanced Open Source Relational Database. https://www.postgresql.org/

    2025

  47. [48]

    Mark Raasveldt and Hannes Mühleisen. 2019. Duckdb: an embeddable analytical database. InProceedings of the 2019 international conference on management of data. 1981–1984

    2019

  48. [49]

    Matthew Russo, Chunwei Liu, Sivaprasad Sudhir, Gerardo Vitagliano, Michael Ca- farella, Tim Kraska, and Samuel Madden. 2025. Abacus: A Cost-Based Optimizer for Semantic Operator Systems.arXiv preprint arXiv:2505.14661(2025)

    work page arXiv 2025

  49. [50]

    salad.com. [n.d.]. Salad.com 5080 Pricing. https://salad.com/pricing. Accessed: 2026-04-25

    2026

  50. [51]

    Nima Shahbazi, Seiji Maekawa, Nikita Bhutani, and Estevam Hruschka. 2026. Om- niTQA: A Cost-Aware System for Hybrid Query Processing over Semi-Structured Data.arXiv preprint arXiv:2604.02444(2026)

    work page internal anchor Pith review Pith/arXiv arXiv 2026

  51. [52]

    Yunxiang Su, Tianjing Zeng, Zhongjun Ding, Yin Lin, Rong Zhu, Zhewei Wei, Bolin Ding, and Jingren Zhou. 2026. Large Language Model-Enhanced Relational Operators: Taxonomy, Benchmark, and Analysis.arXiv preprint arXiv:2603.02537 (2026)

    work page internal anchor Pith review Pith/arXiv arXiv 2026

  52. [53]

    Gemma Team. [n.d.]. Gemma 4. https://deepmind.google/models/gemma/ gemma-4/

  53. [54]

    Gemma Team, Aishwarya Kamath, Johan Ferret, Shreya Pathak, Nino Vieil- lard, Ramona Merhej, Sarah Perrin, Tatiana Matejovicova, Alexandre Ramé, Morgane Rivière, Louis Rouillard, Thomas Mesnard, Geoffrey Cideron, Jean bastien Grill, Sabela Ramos, Edouard Yvinec, Michelle Casbon, Etienne Pot, Ivo Penchev, Gaël Liu, Francesco Visin, Kathleen Kenealy, Lucas B...

    work page internal anchor Pith review Pith/arXiv arXiv 2025

  54. [55]

    Shicheng Liu Jialiang Xu Wesley Tjangnaka, Sina J Semnani Chen Jie Yu, and Monica S Lam. 2024. SUQL: Conversational Search over Structured and Unstruc- tured Data with Large Language Models. (2024)

    2024

  55. [56]

    Transaction Processing Performance Council (TPC). [n.d.]. TPC-DS Benchmark. https://www.tpc.org/tpcds/

  56. [57]

    Transaction Processing Performance Council (TPC). [n.d.]. TPC-H Benchmark. https://www.tpc.org/tpch/

  57. [58]

    Shangqing Tu, Chunyang Li, Jifan Yu, Xiaozhi Wang, Lei Hou, and Juanzi Li

  58. [59]

    Chatlog: Carefully evaluating the evolution of chatgpt across time.arXiv preprint arXiv:2304.14106(2023)

    work page arXiv 2023

  59. [60]

    Vast.ai. [n.d.]. Vast.ai 5080 Pricing. https://cloud.vast.ai. Accessed: 2026-04-25

    2026

  60. [61]

    Darren Yow-Bang Wang, Zhengyuan Shen, Soumya Smruti Mishra, Zhichao Xu, Yifei Teng, and Haibo Ding. 2025. Slot: Structuring the output of large language models.arXiv preprint arXiv:2505.040161, 2 (2025), 3

    work page arXiv 2025

  61. [62]

    Xingyao Wang, Boxuan Li, Yufan Song, Frank F Xu, Xiangru Tang, Mingchen Zhuge, Jiayi Pan, Yueqi Song, Bowen Li, Jaskirat Singh, et al. 2024. Openhands: An open platform for ai software developers as generalist agents.arXiv preprint arXiv:2407.16741(2024)

    work page internal anchor Pith review Pith/arXiv arXiv 2024

  62. [63]

    Johannes Wehrstein, Tiemo Bang, Roman Heinrich, and Carsten Binnig. 2025. GRACEFUL: A Learned Cost Estimator For UDFs.arXiv preprint arXiv:2503.23863 (2025). Parker Glenn and Alfy Samuel

    work page arXiv 2025

  63. [64]

    Brandon T Willard and Rémi Louf. 2023. Efficient guided generation for large language models.arXiv preprint arXiv:2307.09702(2023)

    work page internal anchor Pith review Pith/arXiv arXiv 2023

  64. [65]

    Zhengtong Yan, Gongsheng Yuan, Qingsong Guo, and Jiaheng Lu. 2025. DBMS- LLM Integration Strategies in Industrial and Business Applications: Current Status and Future Challenges.arXiv preprint arXiv:2507.19254(2025)

    work page arXiv 2025

  65. [66]

    Fuheng Zhao, Divyakant Agrawal, and Amr El Abbadi. 2024. Hybrid query- ing over relational databases and large language models.arXiv preprint arXiv:2408.00884(2024)

    work page arXiv 2024

  66. [67]

    $ { model_path }

    Lixi Zhou, Qi Lin, Kanchan Chowdhury, Saif Masood, Alexandre Eichenberger, Hong Min, Alexander Sim, Jie Wang, Yida Wang, Kesheng Wu, et al. 2023. Serving Deep Learning Model in Relational Databases.arXiv preprint arXiv:2310.04696 (2023). 9 APPENDIX 9.1 vLLM Config All local vLLM servers were launched using the following command withvLLM==0.21.0 vllm serve...

    work page arXiv 2023