Sonar-TS: Search-Then-Verify Natural Language Querying for Time Series Databases

Zhao Tan; Yiji Zhao; Shiyu Wang; Chang Xu; Yuxuan Liang; Xiping Liu; Shirui Pan; Ming Jin

cs.AIFeb 19, 2026

Sonar-TS: Search-Then-Verify Natural Language Querying for Time Series Databases

Authors: Zhao Tan, Yiji Zhao, Shiyu Wang, Chang Xu, Yuxuan Liang, Xiping Liu, Shirui Pan, Ming Jin

Organizations: Jiangxi University of Finance and Economics · Griffith University · Yunnan University · Microsoft Research Asia ·

Abstract

Natural Language Querying for Time Series Databases (NLQ4TSDB) aims to assist non-expert users retrieve meaningful events, intervals, and summaries from massive temporal records. However, existing Text-to-SQL methods are not designed for continuous morphological intents such as shapes or anomalies, while time series models struggle to handle ultra-long histories. To address these challenges, we propose Sonar-TS, a neuro-symbolic framework that tackles NLQ4TSDB via a Search-Then-Verify pipeline. Analogous to active sonar, it utilizes a feature index to ping candidate windows via SQL, followed by generated Python programs to lock on and verify candidates against raw signals. To enable effective evaluation, we introduce NLQTSBench, the first large-scale benchmark designed for NLQ over TSDB-scale histories. Our experiments highlight the unique challenges within this domain and demonstrate that Sonar-TS effectively navigates complex temporal queries where traditional methods fail. This work presents the first systematic study of NLQ4TSDB, offering a general framework and evaluation standard to facilitate future research.

Figures & tables

Figure 1: Comparison of querying paradigms. While Text-to-SQL fails to express morphological intents and Time Series Models are limited by context length, Sonar-TS adopts a “Search-Then-Verify” pipeline: it uses SQL to search a symbolic index for candidates and Python to verify them on raw data.

Paradigm / Method	MP	MS	NLG
Text-to-SQL	✗	✓	✓
TSQA Models	✓	✗	✓
TS Similarity Search	✓	✓	✗
Sonar-TS (Ours)	✓	✓	✓

Table 1: Capability comparison of existing paradigms against the fundamental requirements of NLQ4TSDB. Abbreviations: MP (Morphological Primitives), MS (Massive Scalability), NLG (Natural Language Grounding).

Figure 2: The hierarchical taxonomy of tasks in NLQTSBench. The benchmark ranges from Level 1 (Basic Operations) which tests numerical filtering, to Level 2 (Pattern Recognition) for morphological grounding, Level 3 (Semantic Reasoning) for logical composition, and finally Level 4 (Insight Synthesis) for narrative reporting.

Level	Task Sub-Type	Count	Avg. Length
L1: Basic Operations	Atomic Retrieval	216	$\sim$ 2k
L1: Basic Operations	Sliding Window	58	$\sim$ 2k
L2: Pattern Recognition	Shape Identification	129	$\sim$ 0.6k
	Periodicity Detection	120
	Subsequence Matching	111
L3: Semantic Reasoning	Composite Trend	220	$\sim$ 30k

Table 2: Statistics of NLQTSBench. The dataset comprises 1153 queries spanning four complexity levels.

Figure 3: The overview of the Sonar-TS framework. The workflow is organized into three stages: (1) Offline Data Processing constructs compact multi-scale Feature Tables to serve as a queryable index; (2) Online Querying , where the Task Planner and Code Generator synthesize SQL for rapid candidate search and Python for exact verification, guided by domain heuristics distilled from an offline Prompt Cold Start mechanism; and (3) Post-processing translates execution artifacts into a user-friendly interface.

Operator Signature	Mathematical Grounding
detect_period	Autocorrelation Function (ACF) for cyclic behaviors.
find_best_match	Dynamic Time Warping (DTW) for subsequence matching.
detect_changepoints	PELT algorithm for structural segmentation.
calc_trend_slope	Theil-Sen estimator + local slope distribution.
calc_correlation	Pearson correlation for causal links.

Table 3: Representative verification primitives in the operator library.

Figure 4: Snapshot of representative Injected Experiences.

Category	Method	L1: Basic Ops		L2: Pattern Rec.			L3: Semantic Reasoning			L4	Avg.
Category	Method	AR	SW	SI	PD	SM	CT	CxA	CsA	IS	Avg.
Performance on NLQTSBench-Lite (Short Context)
Time Series Models	ChatTS-14B	-	-	0.1768	-	-	0.2431	-	0.1229	-	0.1818
	ITFormer-7B	-	-	0.0736	-	-	0.1500	-	0.1953	-	0.1529
	Time-R1	-	-	0.0320	-	-	0.1395	-	0.1878	-	0.1374
Ours	Sonar-TS	-	-	0.2491	-	-	0.2680	-	0.3615	-	0.3016

Table 4: Main experimental results on NLQTSBench. Abbreviations: AR (Atomic Retrieval), SW (Sliding Window), SI (Shape Ident.), PD (Periodicity Det.), SM (Subseq. Matching), CT (Composite Trend), CxA (Contextual Anomaly), CsA (Causal Anomaly), IS (Insight Synthesis). Best results are bolded , and the best baseline results are underlined .

Variant	L1: Basic		L2: Pattern			L3: Reasoning			L4	Avg.
Variant	AR	SW	SI	PD	SM	CT	CxA	CsA	IS	Avg.
Sonar-TS	0.8609	0.7827	0.3336	0.8625	0.9439	0.2988	0.4767	0.3841	0.7395	0.6144
w/o Self-Correction	0.8302	0.7250	0.3056	0.8618	0.9217	0.2919	0.4813	0.3326	0.7139	0.5930
w/o Feature Tables	0.8658	0.6923	0.1561	0.8804	0.9309	0.0422	0.4968	0.3918	0.7163	0.5430
w/o Experiences	0.8426	0.5591	0.1186	0.8311	0.9138	0.0677	0.2501	0.3612	0.3415	0.4686
w/o Verification	0.8294	0.1870	0.2992	0.0154	0.0396	0.2826	0.0818	0.0592	0.0261	0.2721

Table 5: Ablation study on component contributions. We analyze the impact of removing the Feature Table, Experiences, and the Verification phase. Abbreviations follow Table 4 . Cells highlighted in blue indicate tasks most significantly impacted by the ablation.

Figure 5: Case Study. Text-to-SQL (Left) lacks morphological expressivity, and TS Models (Right) fail the logical constraint. Sonar-TS (Middle) succeeds via Search-Then-Verify.

Appendix figures & tables8 assets

Supplementary material from the paper’s appendix.

Appendix

Figure 6: The human verification interface. Annotators inspect both the global context (top) and the local detail view (bottom, where the injected signal in orange is overlaid on the raw data in blue) to validate the ground truth.

Level	Subtask	Metric
L1	Atomic Retrieval (Global Aggregation)	Relative Accuracy
L1	Atomic Retrieval (Temporal Localization)	Hit Rate
L1	Atomic Retrieval (Interval Discovery)	IoU
L1	Sliding Window	IoU
L2	Shape Identification	IoU
L2	Periodicity Detection	Relative Accuracy

Appendix

Table 6: Task-to-metric lookup for NLQTSBench.

Level	Subtask	#Instances
L2	Shape Identification (SI)	100
L3	Composite Trend (CT)	200
L3	Causal Anomaly (CsA)	200
Total		500

Appendix

Table 7: Task composition of NLQTSBench-Lite (500 instances in total).

Figure 7: Visualization of Multi-Scale SAX Representations. The framework discretizes time series data across hierarchical granularities to support pattern matching at different resolutions: (Top) The Daily View captures high-frequency local fluctuations; (Middle) The Monthly View summarizes intermediate trends; (Bottom) The Yearly View abstracts long-term seasonality. The colored horizontal bars represent the assigned SAX symbols (from level $a$ to $e$ ) overlaid on the raw data (gray line).

Backbone Model	Avg. Score
Qwen3.5-9B	0.3598
Qwen3.5-27B	0.4817
Qwen3.5-35B	0.5359
DeepSeek-V3 (Default)	0.6144

Appendix

Table 8: Sensitivity of Sonar-TS to the LLM backbone.

Alphabet Size ( $\alpha$ )	L2: SI	L3: CT
$\alpha=3$	0.1927	0.2318
$\alpha=5$ (Default)	0.3336	0.2988
$\alpha=7$	0.2896	0.2675

Appendix

Table 9: Sensitivity of morphology tasks to the SAX alphabet size

\alpha

Task Category	Avg. Candidates	Recall
L2: Shape Identification	9.3	0.5482
L3: Composite Trend	5.8	0.4713

Appendix

Table 10: Internal evaluation of the Search stage, reporting average candidates retrieved and recall. Absolute candidate counts depend on benchmark construction details (e.g., the number of injected positives per query and the temporal scope of each instance).

Efficiency Metric	MAC-SQL	Sonar-TS
Offline Construction Time	N/A	$\sim$ 83 s
Online Querying Latency	$\sim$ 15 s	$\sim$ 22 s
Average Token Cost	$\sim$ 8k	$\sim$ 11k
Storage Overhead	N/A	11%

Appendix

Table 11: System overhead and computational cost of Sonar-TS. Time is in seconds (s), token cost is the per-query average, and storage overhead is relative to the raw data size. Reported figures reflect a representative deployment on our experimental setup; concrete values depend on hardware, API latency, and code version.

Time Series QA

Time-series question answering connects natural-language questions to evidence and reasoning over sequential data.

See all 32 papers

TQTS-Bench: A Multi-Syntax Benchmark for Text-to-Query over Time-Series Databases
Sep 28, 2026Fei Lyu, Zhiyi Peng, Jiaming Liu +5LLM Evaluation Time Series QA
College of Computer Science and Electronic Engineering, Hunan University
TT-SAX: A Two-Track Time Series Symbolization for Natural Language Shape Querying
Oct 22, 2026Zhao Tan, Xianglong Bao, Zhipeng Xi +1Time Series QA Time Series Representation Learning
Griffith University, Australia
QuAnTS: Question Answering on Time Series
Nov 7, 2025Felix Divo, Maurice Kraus, Anh Q. Nguyen +5Time Series QA
AI & ML Lab, TU Darmstadt · University of New South Wales · MBZUAI, Abu Dhabi +4

Sep 28, 2026J/K move · Enter open · S save

TQTS-Bench: A Multi-Syntax Benchmark for Text-to-Query over Time-Series Databases

Fei Lyu, Zhiyi Peng, Jiaming Liu, Yixuan Yang+4

College of Computer Science and Electronic Engineering, Hunan University

LLM Evaluation Time Series QA

Open paper AskPDF arXiv

Large language models (LLMs) have significantly advanced natural language querying over relational databases, yet their ability to query time-series databases (TSDBs) remains largely unassessed. Existing benchmarks fail to adequately capture the non-unified query syntaxes, diverse application domains, and unique time-specific query intents inherent to TSDBs. To address this gap, we introduce TQTS-BENCH, a multi-syntax benchmark for evaluating text-to-query capabilities over TSDBs. TQTS-BENCH contains 6,125 high-quality question-answering (QA) pairs spanning 97 TSDBs, 23 distinct query syntaxes, 22 application domains, and 4 types of time-specific query intents. It is constructed through a human-centric AI-assisted workflow, where all QA pairs are carefully reviewed and revised by domain experts to ensure quality and correctness. Extensive evaluations of advanced LLMs and state-of-the-art text-to-query methods reveal challenges in querying TSDBs. Even the best-performing model evaluated, Claude-Opus-5, achieves only 48.98% execution accuracy, while humans reach 87.34%. Error analysis reveals that this performance gap mainly stems from the heterogeneous query syntaxes across different TSDBs, misinterpretation of time-specific intents, and incorrect schema linking. These findings highlight new opportunities to narrow the gap between current LLM capabilities and the requirements of TSDB queries in real-world applications. The benchmark is available at: https://anonymous.4open.science/r/TQTS-Bench-00CD.

Abstract

Time Series QA

Explore similar work

TQTS-Bench: A Multi-Syntax Benchmark for Text-to-Query over Time-Series Databases

TT-SAX: A Two-Track Time Series Symbolization for Natural Language Shape Querying

QuAnTS: Question Answering on Time Series