In a world where data-driven decision-making is critical, businesses are scrambling to find the most efficient ways to extract actionable insights from massive datasets. Uber, a global leader in real-time logistics and transportation, recently shared how their internal Text-to-SQL platform — QueryGPT (If you haven’t checked out the post, check it out here)— is revolutionizing the way their teams interact with data. By enabling employees to simply ask questions in natural language and receive SQL queries in return, Uber has cut query authoring time by 70%. Considering they run about 1.2 million queries per month, this translates into an astonishing 140,000 hours saved monthly.

Yet, this remarkable efficiency isn’t only for tech giants with vast engineering resources. With the advent of open-source solutions like Wren AI, the Text-to-SQL advantage can be democratized. Wren AI Cloud aims to give businesses of all sizes the power to use natural language queries, powered by generative AI, to seamlessly access their data. In this post, we’ll dissect Uber’s QueryGPT from a technical standpoint and highlight how Wren AI Cloud mirrors (and in some cases simplifies) these complex features. The goal is to help you understand how to implement Text-to-SQL in your own operation — no matter your scale.

Understanding Uber’s Technical Approach to Text-to-SQL

Uber’s data platform is a behemoth: it handles trillions of rows, petabytes of data, and millions of queries each month. Traditional SQL authoring is time-consuming and requires users to have strong query-building skills, understand the underlying data models, and know where to find the right tables and columns. QueryGPT removes these roadblocks by using large language models (LLMs) and clever integration into Uber’s existing data ecosystem.

Key Technical Components of designing QueryGPT at Uber, shared in the post of QueryGPT article

1. Workspaces

Workspaces are curated collections of SQL samples and table schemas aligned with specific business domains, such as Mobility, Ads, Core Services, and more. By creating domain-oriented clusters of relevant tables and query templates, QueryGPT narrows the scope of possible data sources, which significantly improves the model’s accuracy when generating SQL.

How It Works: When a user interacts with QueryGPT, the system will first identify which business domain (e.g., Mobility for trips and drivers) is relevant to the query. Within that workspace, QueryGPT will reference a smaller, more focused set of tables and SQL patterns, rather than scanning through Uber’s entire database ecosystem.

Advantages:

• Precision: By limiting the search space to domain-specific references, QueryGPT is more likely to pick the correct tables and columns.
• Reduced Complexity: Users dealing with the Mobility domain, for example, only see tables related to trips, drivers, or documents, which simplifies the data exploration process.
• Customizability: Beyond the System Workspaces that Uber provides by default, users can create their own Custom Workspaces for niche use cases or novel projects that are not covered by standard domains.

2. Intent Agent

After a user inputs a question in natural language, QueryGPT employs an Intent Agent to interpret the user’s intent and determine the most appropriate workspace(s). This step is vital in ensuring the system directs queries to the correct domain and, by extension, the correct subset of tables.

• Intent Detection: Using a large language model, the Intent Agent analyzes the user’s query — looking for keywords, context, and semantics — to map it to one or more domain workspaces. For instance, if the question involves trip data, driver details, or vehicle attributes, the system might route the query to the Mobility workspace.
• Multiple Mappings: Some queries may span multiple domains (e.g., an analysis that touches both Mobility and Ads data), in which case the Intent Agent can map the question to more than one workspace. This ensures that cross-domain queries are also supported.
• Efficiency Gains: Because only the relevant workspaces are considered, QueryGPT reduces the computational overhead of rummaging through irrelevant schemas. This not only boosts accuracy but also shortens the overall query generation time.

3. Table Agent

Once QueryGPT knows which business domain(s) to focus on, the Table Agent proposes a list of specific tables that are most relevant to the user’s request. This step is crucial in large organizations like Uber, where a single domain might contain dozens — or even hundreds — of tables with overlapping or complementary data.

• Table Selection: Drawing on the user’s intent and the curated workspace content, the Table Agent pulls up the most likely candidates needed for the SQL query. It leverages example queries, table relationships, and domain-specific knowledge to decide which tables are pertinent.
• User Verification: The user is then shown a summary of the chosen tables and asked to confirm if they are correct. If something seems off — maybe the system picked an outdated table or missed a new one — the user can edit the list before moving forward. This human-in-the-loop feedback mechanism helps maintain high quality and trustworthiness of the generated SQL.
• Enhanced Collaboration: By allowing users to fine-tune table choices, QueryGPT bridges the gap between automated query generation and domain expertise. Data analysts familiar with specific schemas can quickly ensure QueryGPT is referencing the right data sources.

4. Column Prune Agent

Even with the correct tables identified, large enterprise schemas can contain hundreds of columns, each of which must be described to the language model if there’s a chance it might appear in the query. Such exhaustive detail can hit or exceed token limits during generation, especially when using models like GPT-4 Turbo with high token capacities.

• Pruning Logic: The Column Prune Agent uses an LLM call to filter out columns that are unlikely to be relevant to the user’s question. By doing so, it dramatically reduces the amount of information passed along to the subsequent query generation step.
• Cost and Performance Benefits: With fewer tokens involved, QueryGPT lowers the cost of each LLM call and processes the query faster. Additionally, removing irrelevant columns simplifies the final SQL query, making it more transparent and maintainable.
• Reduced Errors: Handling a slimmer, more targeted schema also reduces the chance of the model selecting the wrong fields — improving both the clarity and the accuracy of the final SQL output.

Bringing It All Together

These four components — Workspaces, the Intent Agent, the Table Agent, and the Column Prune Agent — orchestrate a streamlined, highly efficient text-to-SQL generation process at Uber. By segmenting the solution into domain-specific Workspaces, filtering queries through an Intent Agent, validating table choices via a Table Agent, and pruning unnecessary columns before the final query is generated, QueryGPT ensures high accuracy, cost savings, and quick turnaround. This approach empowers users to interact with Uber’s complex data ecosystem through simple, natural language questions — raising the bar on data accessibility and operational efficiency across the company.

The Direct Business Impacts of QueryGPT at Uber

Technically, QueryGPT is a marvel of LLM integration, prompt engineering, and systems design. The business outcomes reflect this technical mastery:

• 70% Reduction in Query Time: A decrease from about 10 minutes to 3 minutes per query is a massive gain in analyst efficiency.
• 140,000 Hours Saved per Month: This doesn’t just mean cost savings. It enables analysts to spend more time on value-added tasks: interpreting results, optimizing campaigns, improving rider experiences, and strengthening platform reliability.
• Faster Feedback Loops: With quicker access to insights, product managers and data scientists can iterate faster, test hypotheses more frequently, and launch improvements more confidently.
• Competitive Advantage: In an industry where responsiveness to market changes is key, the ability to quickly extract insights from data directly translates to better decision-making and improved customer experiences.

Mapping Uber’s QueryGPT Features to Wren AI Cloud

Uber’s QueryGPT is undoubtedly impressive, but it’s tailored for a massive organization with extensive engineering resources. How can a growing startup, a mid-sized enterprise, or even a solo data practitioner tap into similar technologies? This is where Wren AI comes into play.

Wren AI is an open-source SQL AI agent designed to democratize Text-to-SQL technology. By offering a cloud-based platform that integrates many of the same features as Uber’s QueryGPT, Wren AI aims to level the playing field.

Below is an overview of how Uber’s QueryGPT design aligns with and maps to Wren AI’s features, illustrating how similar principles of workspace separation, intent detection, table selection, and column pruning are implemented in both systems. These parallels demonstrate a shared commitment to delivering a streamlined, secure, and user-friendly text-to-SQL experience.

1. Workspaces → Projects & Organizations in Wren AI

QueryGPT at Uber

In QueryGPT, Workspaces serve as curated collections of SQL samples and table schemas for specific domains, such as Mobility or Core Services. By narrowing the focus to a particular business domain, QueryGPT can more accurately generate SQL queries and ensure that data analysts only interact with context-relevant tables.

Wren AI Equivalent

Wren AI offers similar functionality through its project and organization management features. You can read more about creating organizations and projects in the Wren AI Cloud documentation:

• Create a Project
• Create an Organization

Just like QueryGPT’s Workspaces, Wren AI’s projects let you group and isolate specific data models so that only authorized users can access them. Within a single organization, you can set up multiple projects for different functions or business domains — similar to how QueryGPT sets up separate Workspaces. Access controls in Wren AI ensure that only the right people can view and manage sensitive data, aligning with the same principles of domain-scoped isolation seen in QueryGPT.

Why This Matters

• Targeted Context: By restricting data access to a single workspace or project, the system can better understand user requests and generate more accurate queries.
• Security & Governance: Organizations with sensitive or proprietary data benefit from robust access controls and data policies, ensuring compliance and proper data governance.

2. Intent Agent → Wren AI’s Intent Detection

QueryGPT at Uber

When a user inputs a question, QueryGPT’s Intent Agent identifies which business domain the question belongs to — Mobility, Ads, etc. — and routes the query to the corresponding workspace. This step dramatically narrows the search space for relevant tables and schemas, improving accuracy and speed.

Wren AI Equivalent

Wren AI’s approach to intent detection is described in the Ask documentation. When you ask a question like “How many tables do I have?” or “Explain the customer table to me,” Wren AI automatically discerns whether you’re requesting data retrieval, schema exploration, or if your question falls out of scope (e.g., a casual greeting).

• Data Retrieval Requests: These queries prompt Wren AI to generate SQL, mapping the request to the underlying data schema.
• Schema Exploration: When questions are more about the structure — like listing available tables, describing a table’s columns, or explaining relationships — Wren AI provides in-depth metadata and recommended queries.

Why This Matters

• Automatic Domain Routing: Similar to QueryGPT, Wren AI’s Intent Detection ensures that your request is processed accurately, either leading to SQL generation or a schema exploration response.
• User Guidance: If a request is out-of-scope (e.g., small talk), Wren AI prompts the user to clarify, maintaining a clear focus on data and schema queries.

3. Table Agent → Wren AI’s Table Retrieval Agent

QueryGPT at Uber

In QueryGPT, once the correct domain is identified, the Table Agent proposes which specific tables are necessary to construct the SQL query. Users can either confirm these suggestions or edit them to ensure alignment with their real-world data expertise.

Wren AI Equivalent

In Wren AI, this step corresponds to the table retrieval phase, where we use semantic search to select the top 10 tables most relevant to the user’s question. We look at each table name and its metadata — such as descriptions or tags — to determine the best matches before generating the query.

• Focused Retrieval: By narrowing it down to the most relevant tables, Wren AI saves users from sifting through large, complex schemas.
• Semantic Relevance: Matching a user’s question against the table metadata ensures that the initial query stage is both accurate and efficient.

Why This Matters

• Accuracy: Identifying the right tables up front reduces the likelihood of incorrect or irrelevant queries.
• Efficiency: Semantic search and a concise top-10 table list streamline the retrieval process, saving time for both technical and non-technical users.

4. Column Prune Agent → Wren AI’s Column Pruning

QueryGPT at Uber

‍Large data schemas in enterprise environments can have hundreds of columns per table, potentially hitting token limits when feeding this data into an LLM. The Column Prune Agent filters out unnecessary columns to avoid overloading the model, cutting down on both latency and cost.

‍Wren AI Equivalent

‍Wren AI’s Column Prune Agent serves an identical function: when you connect large numbers of tables and columns, Wren AI prunes those columns that aren’t relevant to the query or the user’s immediate needs. This keeps the system fast, efficient, and affordable to operate at scale.

‍Why This Matters

• Scalability: Reducing token load ensures that the LLM can handle complex queries without timing out or incurring huge processing costs.
• Improved Accuracy: Focusing on the most pertinent columns reduces distractions for the AI, leading to more precise SQL generation.

‍

What’s Even More with Wren AI

‍Beyond its robust text-to-SQL functionality, Wren AI delivers a host of additional features that make data analysis even more convenient, interactive, and accessible for teams of all technical backgrounds.

‍1. Text-to-Chart

‍Wren AI automatically generates insightful charts to visualize your data and uncover meaningful patterns — no additional steps required.

• Automatic Chart Generation: When you pose a question, Wren AI analyzes your dataset and selects the most suitable chart type to display the results.
• Seamless Exploration: Simply switch to the “Chart” tab to see your query results in a bar chart, line chart, pie chart, or another visually compelling format.
• Quick Insights: By representing the data visually, your team can spot trends and correlations faster, making data-driven decisions more intuitive.

Learn more: https://docs.getwren.ai/oss/guide/home/chart

‍2. Data Boilerplates‍

Boilerplates in Wren AI are pre-defined templates designed to simplify your data analysis journey from start to finish.

• Streamlined Setup: Boilerplates come with pre-selected tables and columns, eliminating the need to import large datasets that contain irrelevant information.
• Jumpstart Exploration: Curated sets of frequently asked questions help you quickly discover valuable insights without starting from scratch.
• Supported Integrations: Wren AI already offers boilerplates for HubSpot, GA4, and soon WooCommerce, with more on the way.

Learn more: https://docs.getwren.ai/oss/guide/boilerplates/overview

‍3. Step-by-step SQL breakdown

‍This breakdown walks users through how the AI arrives at certain tables and columns, culminating in the final SQL query.

• User Transparency: Wren AI reveals which tables and columns it intends to use. Users can see the logic behind these choices, enhancing trust and correctness.
• Explainability: Detailed breakdowns help users understand how the AI made its decisions, a critical factor for debugging and compliance in enterprise settings.

Learn more: https://docs.getwren.ai/cloud/guide/home/answer#result-steps

‍4. Connection with Excel and Google Sheets

‍Wren AI makes it easy to share and manipulate query results in familiar spreadsheet tools.

• Excel Add-in: Export data from Wren AI directly into Excel by selecting specific threads or views. This integration reduces manual data copying and speeds up analysis in the environment your team already knows.
• Google Sheets Add-on: Similar functionality for Google’s productivity suite, so teams relying on G Suite can also enjoy seamless data exports.

Learn more:

• Excel Add-in
• Google Sheets Add-on

By incorporating text-to-chart, boilerplates, and deep spreadsheet integrations, Wren AI goes beyond text-to-SQL to deliver a comprehensive, user-friendly ecosystem for data analytics. Whether you need a quick visualization, ready-to-use templates for common data questions, or seamless export to your preferred spreadsheet tool, Wren AI has you covered.

‍Conclusion

‍Both Uber’s QueryGPT and Wren AI share a modular design that solves text-to-SQL challenges by leveraging key steps — workspace or project segmentation, intent detection, table selection, and column pruning. This structured approach ensures accurate, efficient, and scalable SQL generation while maintaining strict data governance and security.

If you’re looking for a powerful, open-source text-to-SQL solution that brings these innovations to your organization, Wren AI offers everything you need to transform how your team interacts with data. With project and organization management, intent detection, step-by-step SQL breakdowns, and column pruning, Wren AI ensures that data is accessible, accurate, and secure — whether you’re a data analyst, engineer, or business leader.

👉 Explore Wren AI’s open-source project on GitHub: https://github.com/Canner/WrenAI

‍👉 Learn more and try Wren AI today: https://getwren.ai/

‍Start simplifying your data workflows and empower your team to get insights faster — no manual SQL writing required.

‍