TLDR: The Datalake Agent is a new system that makes Large Language Models (LLMs) more efficient and cost-effective for translating natural language into SQL (NL2SQL). Instead of feeding LLMs all database information at once, it interactively requests only necessary details, reducing token usage by up to 87% and significantly cutting computational costs, especially for large databases, while maintaining strong performance.
Large Language Models (LLMs) have opened up exciting possibilities for translating natural language queries into SQL (NL2SQL), making it easier for users to interact with databases. However, when dealing with a vast collection of SQL databases, a significant challenge arises: the sheer volume of meta-information (details about database structures and types) that needs to be processed. This often leads to very long prompts for LLMs, resulting in high computational costs and slower processing.
A new research paper, “Agentic NL2SQL to Reduce Computational Costs,” introduces an innovative solution called the Datalake Agent. This agentic system is designed to make LLMs more efficient at solving NL2SQL tasks by intelligently reducing the amount of meta-information they need to process at any given time. Instead of feeding the LLM all database details at once, the Datalake Agent employs an interactive, reasoning-based approach.
How the Datalake Agent Works
The Datalake Agent operates through a structured, iterative process, focusing on acquiring only the necessary information for a specific task. This framework is divided into three core areas:
- Information Acquisition: The LLM starts by gathering general schema knowledge. It uses predefined commands to explore databases, such as retrieving high-level summaries, listing tables within a selected database, and exposing column-level metadata (names and types).
- Iterative Refinement: The LLM follows a hierarchical path, moving from broad data overviews to specific, task-relevant details. Crucially, it can revert to a broader level of information if needed, allowing for flexible, feedback-driven reasoning over complex database structures. This ensures that only truly relevant information is retrieved.
- Query Formulation: Once enough schema information has been gathered, the LLM generates precise SQL queries. The Datalake Agent then executes these queries through a dedicated access layer, designed for scalability and modularity.
Significant Cost Reductions and Improved Efficiency
The researchers evaluated the Datalake Agent against a traditional “direct prompting” strategy, where all schema information is provided upfront. The experiments were conducted on a collection of 23 databases, including real-world and simulated data, totaling up to 319 tables, and involved 100 table question answering tasks.
The results were compelling. The Datalake Agent demonstrated a remarkable reduction in the number of tokens used by the LLM – up to 87% less compared to the direct prompting method. This translates directly into substantial cost savings. For instance, when dealing with 319 tables, the direct solver incurred eight times the cost of the Datalake Agent. For 1000 tasks, this could mean a difference of over $450 with models like OpenAI’s o1.
Beyond cost efficiency, the Datalake Agent also proved to be highly effective in performance. While the direct solver initially performed well on smaller datasets, its accuracy declined sharply as the number of tables increased, especially for more complex queries. In contrast, the Datalake Agent maintained competitive performance, and even improved the LLM’s ability to handle complex queries more effectively in larger and more intricate settings.
Also Read:
- Advancing Conversational Database Interaction with MTSQL-R1’s Agentic Approach
- AI-Powered Workflow Creation: A New Approach for ETL Tools
Addressing Limitations and Future Directions
The study acknowledges a key limitation: the potential for infinite reasoning loops, where the model might repeatedly request the same information without progress. The researchers implemented a mitigation strategy to force an SQL query after 10 repeated requests, ensuring the system doesn’t get stuck. Future work aims to further address this issue, extend evaluations to even larger and more complex datasets, and explore other LLMs and diverse task types to fully understand the Datalake Agent’s potential in real-world applications.
This research marks a significant step towards making NL2SQL with LLMs more practical and affordable for enterprise use cases, particularly in environments with vast and complex data lakes. You can read the full paper here: Agentic NL2SQL to Reduce Computational Costs.
