Discussions with a time-series database:Utilizing large language models to query a time-series database using natural language
Year: 2025
Type: dissertation
Abstract: This thesis explores the development of a natural language interface for querying time-series databases, specifically focusing on ABB Ability™ History, also known as Real-Time Da-tabase. As industrial processes generate large quantities of monitoring data, time-series data-bases have become the primary method for implementing storage. Accessing this data re-quires the user to have specialized knowledge of querying, which creates a barrier for non-technical users. This work investigates approaches for Large Language Models (LLMs) to interpret natural language queries and write the appropriate database requests. As the database is proprietary and training data doesn’t exist, this work excludes approaches requiring fine-tuning. Through a literature review, this work evaluated approaches including Retrieval Augmented Generation (RAG), adaptive retrieval mechanisms, agentic solutions, and LLM frameworks. The analysis determined that an LLM framework-based architecture gives the best balance of control and adaptability when targeting database interaction. A prototype was implemented using the LangGraph library, an approach that allowed for decomposition of the complex problem of query generation into manageable chunks: variable name resolution, creation of data aggregation filters, timestamp interpretation, data fetching and presentation. The prototype demonstrated this approach’s ability to perform all the re-quired functions. Evaluation of the performance showed a high accuracy for most compo-nents. The biggest challenge encountered was resolving the variable name from a long list of potential database identifiers. This thesis provides novel research into effective architectures for database NLIs as well as practical considerations for the development of such systems. The implemented prototype val-idated the chosen approach.
Author: Korpela, Ossi
Language: English
Subfield: Information Systems
Field: Computer Science
Domain: Physical Sciences
Open Access status: green
OpenAlex ID: W7133015816