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

Researchers introduce Sonar-TS, a neuro-symbolic framework that enables natural language querying of time series databases by combining SQL-based feature indexing with Python verification programs. The work addresses limitations in existing Text-to-SQL methods for handling continuous temporal patterns and introduces NLQTSBench, the first large-scale benchmark for evaluating natural language queries against time series data at scale.

Sonar-TS represents a meaningful advancement in making time series data more accessible to non-technical users through natural language interfaces. The framework tackles a genuine technical gap: existing Text-to-SQL systems excel at structured queries but struggle with morphological intents like shape recognition or anomaly detection, while time series models fail at ultra-long history processing. The Search-Then-Verify pipeline mirrors sonar technology by first identifying candidate windows through SQL queries against feature indices, then validating them against raw signals using generated Python code.

This work emerges from the broader trend of democratizing data access through language models. As organizations accumulate massive temporal datasets in monitoring systems, financial records, and sensor networks, the ability to query these repositories conversationally addresses a real operational bottleneck. Traditional approaches force users to understand both time series specifics and database syntax.

The introduction of NLQTSBench signals growing maturity in the field—establishing standardized evaluation benchmarks typically precedes widespread adoption and funding. For database vendors and data analytics platforms, this framework offers a potential competitive differentiator. For developers, Sonar-TS provides an open methodology for integrating language models with time series systems. The neuro-symbolic approach (combining learned and rule-based components) has proven more robust than pure deep learning in similar domains.

The immediate impact remains academic, but the framework could influence enterprise tools within 18-24 months. Watch for adoption by monitoring platforms like Datadog or New Relic, and whether cloud providers integrate similar capabilities. The benchmark's quality will determine whether this becomes a reference standard or niche contribution.

• →Sonar-TS combines SQL indexing with Python verification to query time series for both structured events and morphological patterns like anomalies
• →NLQTSBench provides the first standardized evaluation benchmark for natural language querying against time series databases at scale
• →The neuro-symbolic Search-Then-Verify approach outperforms traditional Text-to-SQL and pure time series models on complex temporal queries
• →Framework addresses the gap between non-expert user needs and existing database query complexity in monitoring and analytics platforms
• →Establishes methodology for future research in natural language interfaces for temporal data systems across enterprise applications