Datatune tackles a growing gap in data‑intelligence tools. While Text‑to‑SQL solutions have dominated, LLMs struggle to ingest raw tables because their context length caps far below the billions of tokens a typical database holds. The result is a mismatch between data volume and model capacity.

By giving models full access to the data, Datatune breaks the abstraction barrier. It streams rows in batch mode, sending each chunk to the LLM while Dask splits the workload across partitions. This parallelism lets a single prompt transform millions of records without exceeding the model’s token window.

Datatune exposes four core primitives—MAP, FILTER, EXPAND, and REDUCE—that can be invoked with natural‑language prompts. An Agent layer lets users chain these steps in a single sentence, even injecting Python code for numeric calculations. The result is a seamless, code‑free workflow for complex queries.

Built as an open‑source Python package, Datatune integrates with Ibis, DuckDB, Postgres, and MySQL. The project invites community contributions on GitHub, promising rapid evolution as more data sources and LLM models join the ecosystem. For analysts, it offers a practical bridge between raw tables and semantic AI.