Authors: Matthias Urban, Carsten Binnig

Paper accepted to VLDB'25: “ELEET: Efficient Learned Query Execution over Text and Tables”

Data systems increasingly support LLM calls to process text in databases. But are LLMs the best solution in terms of accuracy and cost? We introduce ELEET, a query execution engine that integrates text sources into databases, allowing users to query texts seamlessly as if they were tables. Instead of relying on expensive LLMs, ELEET leverages a targeted small language model (SLM) to efficiently extract structured data from text. Our evaluation shows that ELEET achieves up to 575× speedups over LLM-based approaches without sacrificing accuracy.

• LLMs in Databases: Modern data systems increasingly support LLM calls to process text. But are they the best solution to process large amounts of texts?
• ELEET: We introduce ELEET, which allows users to query text collections in databases seamlessly as if they were tables. However, we do not rely on LLMs.
• SLM-Based Approach: Instead of expensive LLMs, ELEET uses a specialized small language model (SLM) for efficient structured data extraction from texts.
• Performance Gains: ELEET achieves up to 575× speedups over LLM-based approaches without sacrificing accuracy.

The paper can be downloaded here (opens in new tab).