LLMs for industrial databases: an agro-food production plant use case

Area of Science:

• Industrial data management
• Natural Language Processing
• Database query optimization

Background:

• Industrial time-series databases (e.g., InfluxDB 2.0) require specialized query languages (InfluxQL, Flux).
• Domain experts often lack proficiency in these query languages.
• A lack of labeled corpora hinders natural language to query translation.

Purpose of the Study:

• To develop a system for translating natural language questions into InfluxQL and Flux queries.
• To enable domain experts to extract value from industrial time-series data without specialized query language knowledge.
• To create a robust NL-to-query generation pipeline for InfluxDB.

Main Methods:

• A modular system incorporating a knowledge graph, entity-linking, a fine-tuned language model, and query validation.
• Development of an automated synthetic dataset distillation pipeline using a teacher model and InfluxDB documentation.
• Parameter-efficient fine-tuning of compact Small Language Models (SLMs) using domain-level conditioning and instruction tuning.

Main Results:

• The system reliably generates syntactically valid queries.
• Functional evaluation showed high parser validity, moderate execution success, and lower result correctness, indicating a semantic layer gap.
• Compact SLMs were successfully fine-tuned for NL-to-query generation.

Conclusions:

• A complete NL-to-InfluxDB pipeline was developed, grounded in a semantic representation of an industrial schema.
• A documentation-driven synthetic data generation process with automatic verification was established.
• A parameter-efficient fine-tuning strategy enables query generation with lightweight models for resource-constrained environments.