Spider 2.0: Evaluating Language Models on Real-World Enterprise Text-to-SQL Workflows

Real-world enterprise text-to-SQL workflows often involve complex cloud or local data across various database systems, multiple SQL queries in various dialects, and diverse operations from data transformation to analytics. We introduce Spider 2.0, an evaluation framework comprising 632 real-world text-to-SQL workflow problems derived from enterprise-level database use cases. The databases in Spider 2.0 are sourced from real data applications, often containing over 1,000 columns and stored in local or cloud database systems such as BigQuery and Snowflake. We show that solving problems in Spider 2.0 frequently requires understanding and searching through database metadata, dialect documentation, and even project-level codebases. This challenge calls for models to interact with complex SQL workflow environments, process extremely long contexts, perform intricate reasoning, and generate multiple SQL queries with diverse operations, often exceeding 100 lines, which goes far beyond traditional text-to-SQL challenges. Our evaluations indicate that based on o1-preview, our code agent framework successfully solves only 21.3% of the tasks, compared with 91.2% on Spider 1.0 and 73.0% on BIRD. Our results on Spider 2.0 show that while language models have demonstrated remarkable performance in code generation -especially in prior text-to-SQL benchmarks -they require significant improvement in order to achieve adequate performance for real-world enterprise usage. Progress on Spider 2.0 represents crucial steps towards developing intelligent, autonomous, code agents for real-world enterprise settings. Our code, baseline models, and data are available at spider2-sql.github.io.

Unlike previous datasets, Spider 2.0's agentic task setting does not rely on pre-prepared inputs (question and database schema) or expected outputs (predicted SQL). Instead, it incorporates a real project codebase and a database interface. This complexity extends beyond merely predicting an SQL query; it involves navigating the project and dynamically interacting with complex databases through SQL queries and command-line scripts (in Python or Shell). The task objective is to perform intricate data transformations within the database or to extract analytical insights from the data. This task setting closely mirrors real-world enterprise SQL workflows, requiring the model to refer to the codebase and documentation, generate multiple SQL queries, and dynamically interact with the environment to complete complex tasks and derive the final result. To simplify performance comparisons with previous text-to-SQL methods and benchmarks, and to support faster development and evaluation, we also introduce Spider 2.0-lite and Spider 2.0-snow, self-contained datasets with preprocessed database schema and documentation, the former is hosted on BigQuery, Snowflake, and SQLite, while the latter is entirely hosted on Snowflake. This setting omits the codebase and restricts output to SQL only, thus eliminating the need to predict final answers or transform the database. While they are sourced from the same raw data as Spider 2.0, these settings are not easier than Spider 2.0 because text-to-SQL setting have access to less information (e.g., execution feedback). We present Spider 2.0-lite and Spider 2.0-snow as direct text-input-to-SQL-output challenges that are easier to work with using current advanced text-to-SQL parsers, and Spider 2.0 as