Skip to content

Open Source

IntentQL is an open-source semantic compiler for natural-language analytics over Postgres. The compiler is public so users can inspect, test, extend, and evaluate the full path from semantic hints to parameterized SQL.

License

IntentQL is licensed under the Apache License 2.0. The repository's LICENSE file is the authoritative license text.

The BIRD benchmark dataset is maintained by its own authors and uses its own license. IntentQL does not bundle downloaded BIRD databases. Follow BIRD's license and attribution requirements when downloading or redistributing its benchmark material.

Public Project Scope

The open-source project includes the capabilities that define IntentQL:

  • QueryIntent and QueryPlan specifications;
  • schema introspection and schema configuration;
  • evidence and value resolution;
  • intent normalization;
  • semantic planning and join planning;
  • validation and semantic linting;
  • deterministic SQL compilation;
  • parameterized execution;
  • model adapters;
  • benchmark runners and generic regression tests.

Users can run IntentQL locally, connect their own model and Postgres database, and evaluate the resulting plans and SQL without using an IntentQL-operated service.

Governance

IntentQL is led and maintained by Alexander Abakah.

Contributors may propose changes through issues and pull requests. A contribution becomes part of the official project only after it is reviewed and merged by the lead maintainer. Only the lead maintainer publishes official releases, creates official release tags, and publishes the official intentql package to PyPI.

The Apache License 2.0 permits forks and redistribution under its terms. Forks and modified distributions must not present themselves as official IntentQL releases.

Contribution Standard

IntentQL welcomes contributions that improve general semantic reasoning, compiler expressiveness, validation, safety, integrations, documentation, and evaluation.

The central technical rule is:

A benchmark failure may reveal a missing general capability, but benchmark-specific behavior does not belong in the compiler.

Good contributions explain the reusable capability being added and include domain-neutral tests. For example, support for ranked aggregate comparisons is reusable; recognizing a specific entity from a benchmark database is not.

See Contributing for the development workflow and Benchmarks for the evaluation methodology.