DataGallery Builds Data Foundation for Enterprise AI Agents

DataGallery is an open-source data-agent platform for enterprises that need AI systems to work with real business data – not just demonstrations.

As organisations move AI from pilots into production, data assets, business semantics, domain knowledge and execution tools often remain scattered across systems, with inconsistent quality and limited context. DataGallery brings these elements together in a governed foundation, enabling agents to retrieve information, reason over data, call tools and evaluate results within enterprise data environments.

For business and technology leaders, the goal is straightforward: make enterprise data usable by AI while maintaining governance, traceability and operational control.

DataGallery’s enterprise focus is reflected in both benchmark performance and production deployments.

In June, DataGallery-Text2SQL achieved 77.53% execution accuracy on the BIRD benchmark, which evaluates Text-to-SQL performance across complex, real-world databases. It ranked fourth on the public leaderboard and first among open-source solutions.

At a leading bank, DataGallery provides natural-language data access and agent-assisted analytics to approximately 20,000 data analysts and 200,000 management and marketing users.

The deployment illustrates how governed enterprise data, business semantics and domain knowledge can be turned into practical AI-assisted workflows.

Enterprise agents need more than a capable model. They must identify the correct data source, interpret business metrics, respect permission boundaries and provide traceable evidence. DataGallery addresses these as system-level requirements, bringing together data assets, tool execution, business semantics, knowledge and evaluation across the data-agent lifecycle.

Its architecture centres on three capabilities:

• DataAgent translates user intent into controlled execution across tools and data systems.
• The Unified Semantic Engine provides a governed map of enterprise data, including business definitions, relationships and permissions. 
• KnowEdge turns long-form documents and expert knowledge into structured, retrievable assets for AI systems.

Together, these components provide a foundation for enterprise AI applications across analytics, knowledge work, research and industry-specific workflows.

DataAgent turns business context into action

DataAgent is DataGallery’s orchestration layer for enterprise data work. It translates user intent into a controlled sequence of tasks, including source selection, metric interpretation, tool execution, result validation and output generation.

This allows agents to handle multi-step workflows rather than isolated questions. DataAgent supports data analysis, feature engineering and knowledge-based construction while operating within defined permission, cost and safety controls. DataAgent therefore functions as an execution layer for repeatable data work, with DataGallery providing the context, guardrails and evaluation required for production use.

Unified Semantic Engine gives agents trusted data context

The Unified Semantic Engine connects structured, semi-structured and unstructured data from heterogeneous sources and organises it into a governed semantic layer. This includes tables, columns, relationships, metric definitions, entities, business terms and permissions.

Instead of giving agents direct access to a complex data warehouse, it provides a business-aligned map of how the organisation understands its data. This is particularly important in large environments, where a model cannot inspect every table or column for each query.

Business questions about revenue, conversion, active users, risk or churn rarely correspond to a single database field. The engine makes these definitions explicit and uses schema linking to identify the most relevant tables, columns and relationships. This narrows the search space and improves the reliability of natural-language data access at scale.

KnowEdge makes enterprise knowledge usable by AI

KnowEdge converts long-form documents into structured, citation-ready knowledge assets. It processes financial reports, research papers, policy documents and internal materials into document trees containing hierarchy, page anchors, section context, summaries and retrieval-optimised chunks.

Agents can search across complex documents, assemble evidence, cite sources and reason over domain-specific material with reduced manual review. The practical impact includes less repetitive document review, improved traceability of AI-generated answers and reuse of institutional knowledge that would otherwise remain locked in files and expert workflows.

Combined with DataAgent and the Unified Semantic Engine, KnowEdge extends DataGallery’s scope beyond Text-to-SQL and document Q&A.

How DataGallery performs in enterprise scenarios

DataGallery’s BIRD performance is supported by a pipeline covering metadata preparation, schema linking, multi-path SQL generation, execution verification and confidence-based selection. Its schema-linking component achieved 99.1% micro recall and a 95% success rate on the evaluated set.

The same approach has been applied in an enterprise data environment containing more than 3,000 tables, where DataGallery was used to build a semantic layer and support complex SQL execution across a large schema. The deployment shows that reliable natural-language access depends not only on model performance, but also on governed and accessible business semantics.

DataGallery also supports document-heavy knowledge work. Under the project’s evaluation protocol, KnowEdge achieved 97.33% accuracy on the public 150-question FinanceBench sample. In AI-for-Science applications, the platform has been used to screen, process and analyse approximately 80 million scientific papers.

Across these scenarios, DataGallery provides agents with a governed view of databases, documents and business workflows. Its role is to help enterprises move AI from isolated pilots into daily data work while maintaining control over accuracy, cost and accountability.

To learn more about DataGallery and explore its open-source projects, visit:

GitCode: https://gitcode.com/datagallery | GitHub: https://github.com/datagallery-ai