AI-Native Database Query Interface

Part of the worlds-biggest-software-project initiative.

An open, AI-native natural-language-to-SQL layer with schema-aware query generation, plain-English explanations, and optimisation suggestions across major database dialects.

This project is a model-agnostic NL-to-SQL interface for developers, analysts, and data platform teams who need self-serve access to relational data without writing SQL by hand. It combines live schema introspection, semantic-layer-aware translation, and query explanation so non-experts can query data confidently and engineers can audit what the AI generates.

Why AI-Native Database Query Interface?

Standalone NL-to-SQL tools (AI2SQL, BlazeSQL, Text2SQL.ai) are proprietary SaaS at USD 15-29/user/month and lack a first-class semantic layer, so accuracy degrades on poorly named enterprise schemas.
Hyperscaler offerings (Snowflake Cortex Analyst, BigQuery Gemini, Databricks Assistant, Microsoft Copilot, Oracle SQL AI) are locked to a single platform and cannot federate queries across heterogeneous databases.
Open-source alternatives are fragmented: Vanna.ai requires cold-start training, DBHub has no semantic layer or multi-user isolation, and Wren AI's MDL semantic model must be authored manually.
JetBrains DataGrip's AI Assistant is desktop-only and cannot be embedded as a service in other applications.
Enterprise needs - per-user row-level security in the NL layer, query audit trails, and prompt-injection defence - are weak or absent across the surveyed tools.

Key Features

Natural Language to SQL Core

Natural language input translated to valid SQL for PostgreSQL, MySQL, and Snowflake at minimum, with extensibility to BigQuery, SQL Server, Redshift, Oracle, SQLite, and DuckDB.
Live schema introspection so generated queries reference real table and column names rather than guessed identifiers.
Plain-English query explanation describing what the generated SQL does, clause by clause.
Read-only safety mode that blocks INSERT, UPDATE, DELETE, and DDL until a user explicitly permits write operations.

Semantic Layer and Accuracy

YAML-based semantic layer for metric definitions, synonyms, and table relationships, compatible with the dbt MetricFlow format.
Multi-turn conversational refinement so users can iteratively narrow results ("now just show Q1", "exclude internal users").
Backlog goal of auto-generating the semantic layer from DDL, query history, and schema comments to reduce upfront curation cost.

Enterprise Controls

Per-user row-level security applied at the NL translation layer rather than delegated to the downstream database.
Immutable query audit log capturing each natural-language question, the generated SQL, and the result for compliance and debugging.
Streaming API responses for low-latency chat UX.
Backlog goal of prompt-injection detection on NL inputs.

Integration Surface

MCP (Model Context Protocol) server interface so the tool works with any MCP-compatible AI assistant (Claude, GitHub Copilot, Cursor, VS Code, Codex).
REST API for embedding NL-to-SQL into third-party applications.
Backlog goal of cross-database federated NL queries that decompose a single question into sub-queries across Snowflake, BigQuery, and Postgres and join the results.
Backlog goal of basic text-to-chart output alongside generated SQL.

AI-Native Advantage

Rather than using AI only as a one-shot prompt-to-SQL translator, this project treats AI as the substrate of the whole query pipeline: auto-generating semantic layer artefacts from DDL and query history, explaining and rewriting slow queries, refining results across multi-turn dialogue, and routing federated queries across heterogeneous databases. Combined with row-level security applied inside the NL layer, this addresses the "messy enterprise database" gap where current tools drop to 50-70% accuracy and where audit, security, and federation are weak across incumbents.

Tech Stack & Deployment

Deployment is intended to be self-hosted first (Docker) with an optional managed mode, exposing an MCP server for AI-assistant integration and a REST API for embedding. The project aligns with the SQL:2023 standard for dialect generation, MCP for AI client interoperability, JDBC/ODBC for schema introspection and execution, and the dbt semantic layer format for metric definitions. The architecture is LLM-agnostic, following the pattern set by Vanna.ai and Wren AI of pluggable model backends (OpenAI, Anthropic, Gemini, Ollama, Bedrock).

Market Context

NL-to-SQL sits at the intersection of the broader BI market (USD 33 billion in 2024, growing at ~8% CAGR) and conversational AI, with Gartner projecting that 80% of organisations will adopt AI-driven data solutions by 2026. Standalone tools price at USD 15-29/user/month for professional tiers, while hyperscalers bundle NL-to-SQL into existing database subscriptions. Primary buyers are business analysts and product managers seeking self-serve data access, data engineers building semantic layers, enterprise data platform teams standardising NL across multiple databases, and CISOs requiring auditing and access control on natural-language data requests.

Project Status

This project is in the research and specification phase.
Contributions, feedback, and domain expertise are welcome.

Contributing

We welcome contributions from developers, domain experts, and potential users. See CONTRIBUTING.md for guidelines.

Important: All contributions must be your own original work or clearly attributed open-source material with a compatible licence. Copyright infringement and licence violations will not be tolerated and will result in immediate removal of the offending contribution. If you are unsure whether a piece of code, text, or other material is safe to contribute, open an issue and ask before submitting.

Licence

Licence to be determined. See discussion for context.