Skip to content

Project Roadmap

IntentQL's goal is to make natural-language analytics more reliable by moving structural reasoning from model-generated SQL into a typed semantic compiler.

The roadmap is organized around capabilities, not benchmark-specific cases.

Now: Broaden Semantic Coverage

  • resolve tables, columns, values, and joins from schema evidence;
  • improve ranking, extrema, comparisons, ranges, and temporal reasoning;
  • support richer aggregate formulas and distinct-count semantics;
  • make unsupported requests fail clearly instead of producing plausible wrong SQL;
  • protect previous behavior with domain-neutral regression tests.

Next: Make Evaluation Complete

  • run a clean evaluation over all 500 BIRD Mini-Dev cases;
  • publish model, configuration, commit SHA, and reproducible commands;
  • classify failures by compiler, resolver, model, ambiguity, and infrastructure;
  • add more independent datasets and real-world schemas;
  • measure accuracy, latency, model cost, and clarification rate separately.

Next: Improve Developer Experience

  • simplify first-run schema generation and review;
  • provide clearer plan traces and resolution explanations;
  • improve adapter support for local and inexpensive models;
  • add examples for embedding IntentQL in APIs and applications;
  • document extension interfaces for custom schema metadata and resolvers.

Later: Production And Enterprise Operations

  • native asynchronous execution support;
  • richer policy enforcement and audit artifacts;
  • evaluation and regression dashboards;
  • managed and private deployment options;
  • authentication, organizations, RBAC, and enterprise identity;
  • connectors beyond Postgres where the compiler's guarantees can be preserved.

Definition Of Progress

IntentQL is improving when:

  1. more useful questions can be represented by the IR;
  2. fewer questions require a capable or expensive model;
  3. failures become explicit instead of silently wrong;
  4. generic improvements transfer across unrelated schemas;
  5. users can reproduce the claims we publish.

The implementation-level roadmap also lives in the main repository's ROADMAP.md.