CADENZA in Action Breaking the Monolith with Intent-Dependent Plan Spaces for Semantic Queries

Jul 16, 2026

Semantic query processing engines execute semantic operators whose behavior is defined by natural-language intents and carried out through model inference over multimodal data. Existing optimizers mainly tune monolithic operator implementations, creating a costly trade-off between expensive model calls and cheaper methods that miss intent-dependent semantics. CADENZA compiles each intent into decomposed steps, selects physical implementations, and optimizes parameters according to user quality–latency–cost preferences. The web demo explores plan decomposition, plan optimization, and preference-driven selection across multimodal databases.

LanguageEnglish

CategoryTechnology

FormatPDF

Pages4

Reading Time10 min

Introduction

Semantic query processing and intent-based operators

Limitations of monolithic optimizers

CADENZA: intent decomposition and plan optimization

CADENZA in Action: Breaking the Monolith with Intent-Dependent Plan Spaces for Semantic Queries Jaehyun Ha [jhha@dblab.postech.ac.kr](jhha@dblab.postech.ac.kr)[ ](jhha@dblab.postech.ac.kr)GSAI, POSTECH Pohang, Korea Yongjoo Park [yongjoo@illinois.edu](yongjoo@illinois.edu)[ ](yongjoo@illinois.edu)Univ. of Illinois Urbana-Champaign Urbana, USA Wook-Shin Han∗ [wshan@dblab.postech.ac.kr](wshan@dblab.postech.ac.kr)[ ](wshan@dblab.postech.ac.kr)GSAI, POSTECH Pohang, Korea arXiv :2607 .0 1468v 1 [ cs .DB] 1 Jul 2026 Abstract Semantic query processing engines execute semantic operators, whose behavior is specified by natural-language intents, via model inference over multimodal data. Most existing optimizers optimize the operators at the granularity of monolithic implementations—such as LLMs and embedding models—forcing a trade-off between expensive model calls and cheaper alternatives that fail to capture intent-dependent semantics. We present CADENZA, a semantic operator optimizer that compiles an intent into decomposed steps, selects concrete physical implementations for each step, and tunes their parameters under user-specified quality–latency–cost preferences. In this demonstration, users interact with CADENZA through a web interface over multimodal databases, exploring how an intent is decomposed into alternative plans, how each plan is optimized, and how different preferences yield different winning plans. PVLDB Reference Format: Jaehyun Ha, Yongjoo Park, and Wook-Shin Han. CADENZA in Action: Breaking the Monolith with Intent-Dependent Plan Spaces for Semantic Queries. PVLDB, 19(12): XXX-XXX, 2026 . doi:XX.XX/XXX.XX PVLDB Artifact Availability: The source code, data, and/or other artifacts have been made available at [https://github.com/postechdblab/CADENZA](https://github.com/postechdblab/CADENZA). 1 Introduction With the advent of Large Language Models (LLMs), a new class of semantic query processing engines (SQPEs) has emerged [4, 5, 8, 10, 11] . Semantic operators—such as SemFilter, SemJoin, and SemMap—are operators whose behaviors are specified by natural-language task descriptions (i.e., intents) and executed via model inference over multimodal data (e.g., text and images). These operators simplify complex multimodal analysis. For example, suppose an e-commerce platform onboards a vendor who delivers product descriptions (text with brand, category, specs, etc.) and product photos (images with logos but no metadata) as databases with no shared identifier. To build a unified catalog, a data engineer can execute a SemJoin to“match product descriptions with product photos of the same brand.” Most existing semantic query optimizers [4, 5, 8–11] optimize at the granularity of a monolithic operator implementation—e.g., ∗ Corresponding author. This work is licensed under the Creative Commons BY-NC-ND 4.0 International License. Visit [https://creativecommons.org/licenses/by-nc-nd/4.0/ to view a copy of](https://creativecommons.org/licenses/by-nc-nd/4.0/ to view a copy of)[ ](https://creativecommons.org/licenses/by-nc-nd/4.0/ to view a copy of)[this license. For any use beyond those covered by this license](this license. For any use beyond those covered by this license), [obtain permission by](obtain permission by)[emailing info@vldb.org. Copyright](emailing info@vldb.org. Copyright) is held by the owner/author(s). Publication rights licensed to the VLDB Endowment. Proceedings of the VLDB Endowment, Vol. 19, No. 12 ISSN 2150-8097 . doi:XX.XX/XXX.XX (C) | SCAN(images) | image | ImgObj | object | 􀀡5+9$4! %)5:5 | | --- | --- | --- | --- | --- | | | | | | | | SCAN(desc) | desc | TxtNER | entity, | 􀀡!"\#$ %&'( | | --- | --- | --- | --- | --- | | | | | type | | | TxtImgSim | score | 􀀡3456$78 | | --- | --- | --- | | | | | Figure 1: Three candidate logical plans for SemJoin with“match product descriptions with product photos of the same brand.” Each node is an operator and each

FAQs

How are different execution plans selected for a user’s preferences?

CADENZA optimizes parameters under user-specified quality–latency–cost preferences, leading to different winning plans among alternative decompositions for the same intent.

Jul 16, 2026

Upload by Lucas Martin