QueryGPT

QueryGPT is an internal AI-powered natural-language-to-SQL system developed by Uber to help employees generate complex SQL queries from plain-English questions. It was created to reduce the time and technical knowledge required to access Uber’s large-scale data environment, particularly for engineers, operations managers, data scientists, and other teams that regularly work with business data.

QueryGPT combines large language models (LLMs), retrieval-augmented generation (RAG), vector databases, similarity search, business-domain context, SQL examples, and database schemas. Instead of requiring users to manually search through data documentation and construct SQL queries from scratch, the system interprets a natural-language request and uses relevant schemas and previously created SQL examples to generate a query. Uber reported that a typical query could take around 10 minutes to author manually, while QueryGPT could generate a sufficiently reliable query in about three minutes.

The project originated during Uber’s Generative AI Hackdays in May 2023. It subsequently went through more than 20 iterations as Uber improved the retrieval, query-generation, evaluation, and user-feedback components. By September 2024, Uber described QueryGPT as a production-ready service running within its data platform.

One of QueryGPT’s central capabilities is natural-language SQL generation. Users can describe the information they need using ordinary English, and the system generates SQL together with an explanation of how the query was produced. This makes database querying more accessible to users who understand the business question but may not have extensive SQL knowledge.

QueryGPT uses specialized workspaces to improve the relevance of generated queries. These workspaces contain curated collections of SQL samples and database tables associated with particular business domains. Uber’s system includes workspaces covering areas such as Mobility, Ads, Core Services, Platform Engineering, and IT. Users can also create custom workspaces when an existing system workspace does not match their requirements.

The system uses an Intent Agent as an intermediate reasoning step. Instead of sending a broad user question directly to the SQL-generation model, the Intent Agent identifies the relevant business domain and maps the question to appropriate workspaces. This reduces the search space and helps the system locate more relevant schemas and SQL examples.

QueryGPT also includes a Table Agent. This component identifies the database tables that may be required to answer a question and allows the user to review or modify the proposed table selection before SQL generation. This human-in-the-loop approach was introduced after Uber observed that automatically selected tables were sometimes incorrect.

Another component is the Column Prune Agent. Large enterprise tables can contain hundreds of columns, creating extremely large prompts for an LLM. QueryGPT uses the Column Prune Agent to remove irrelevant columns from the schemas before sending them to the SQL-generation model. Uber reported that this reduced token usage, cost, and latency while helping the system work with large schemas.

The platform also includes a structured evaluation system. Uber created a set of verified question-to-SQL examples and evaluates QueryGPT on intent accuracy, table selection, successful query execution, whether the query returns useful output, and qualitative similarity between generated SQL and verified SQL. This allows the team to identify regressions and measure improvements over time.

QueryGPT uses OpenAI models as part of its architecture. Uber’s 2024 description specifically states that the production system used GPT-4 Turbo with a 128K-token limit and applied additional agents and retrieval techniques to control the amount of schema information sent to the model.

The system is designed around Uber’s internal data ecosystem rather than generic public databases. Its knowledge of Uber-specific business terminology, schemas, SQL patterns, and domain concepts allows it to generate queries that are tailored to Uber’s data environment.

QueryGPT also supports an iterative chat-style workflow. Users can refine generated queries and provide feedback instead of treating the first generated query as final. Uber has experimented with additional validation mechanisms to reduce hallucinations and improve reliability.

The system is not completely free from AI limitations. Uber notes that LLM hallucinations can still produce nonexistent tables or columns, while short or poorly contextualized user prompts can make query generation more difficult. Large and inconsistently documented schemas also remain challenging. Uber therefore emphasizes evaluation, human feedback, iterative improvements, and appropriate targeting of user groups.

By September 2024, QueryGPT had been released on a limited basis to some Operations and Support teams. Uber reported approximately 300 daily active users in that limited release, with around 78% of surveyed users saying that the generated queries reduced the time they would otherwise have spent writing SQL manually.

QueryGPT is therefore best understood as an enterprise internal data-access and SQL-generation system rather than a general-purpose public AI chatbot. Its main value comes from combining generative AI with structured enterprise knowledge, database schemas, SQL examples, business-domain context, human review, and systematic evaluation.

Pricing

QueryGPT does not have publicly listed commercial pricing. It was developed by Uber as an internal data-platform service and was described as being released to selected Uber Operations and Support teams rather than offered as a public SaaS product.

Review

QueryGPT demonstrates how generative AI can be integrated into a large enterprise data environment rather than used only as a general conversational assistant. Its strongest feature is the combination of natural-language interaction with Uber-specific database knowledge, SQL examples, business terminology, and schema information.

The multi-agent architecture is another important strength. The Intent Agent, Table Agent, and Column Prune Agent each handle specialized parts of the query-generation process, while human users can review table selections before the final SQL is generated. This provides more control than simply sending a question to a general-purpose LLM.

QueryGPT’s evaluation framework is also a significant advantage. Uber evaluates intent selection, table overlap, query execution, returned output, and similarity to verified SQL. This provides a structured way to monitor the quality of an AI system that directly affects data analysis.

The main limitation is that QueryGPT is closely tied to Uber’s internal data ecosystem. It is not presented as a public application that organizations can independently sign up for and connect to their own databases. Its architecture also demonstrates the continuing challenges of enterprise text-to-SQL systems, including hallucinated tables or columns, incomplete user prompts, very large schemas, and differences between technically valid SQL and the SQL that correctly answers a business question.

For an AI tool directory, QueryGPT is best categorized as an enterprise AI data-analysis and natural-language SQL system. It is particularly notable as an example of production-scale text-to-SQL architecture using retrieval, specialized agents, human-in-the-loop validation, and continuous evaluation.