Querying an OLAP Cube in Natural Language: Comparison of Querying with LLMs Without and With the Use of a Vector Database

This bachelor thesis focuses on querying OLAP cubes using natural language with the support of large language models. The main objective is to compare two approaches to identifying rele- vant dimensions and measures required for constructing analytical queries. The first approach relies on providing the full metadata context to the language model without prior filtering, while the second approach utilizes vector-based retrieval to preselect relevant metadata. An experimental prototype was designed and implemented to process natural language que- ries, select relevant metadata, generate corresponding MDX queries, and return final answers. The experiments were conducted on an OLAP cube built on the Adventure Works 2019 da- tabase. The evaluation was based on the correctness of metadata selection, correctness of the final answer, and processing time. The results of the conducted experiment indicate that the retrieval-based approach provides a more favorable trade-off between result quality and processing time, especially under higher system load. The results indicate that the way context is provided to the language model may significantly affect the practical usability of the solution.