Ask your trade question
Plain English in, DuckDB SQL and a result table out. The model writes the query; the query runs against the same parquet corpus as every other page on this site. Numbers you see are computed by DuckDB, not generated by the language model.
What AI does and does not do here
- Model: Anthropic
claude-haiku-4-5via the public Anthropic API, prompted with the schema of the relevant parquet tables. - AI-written: the DuckDB SQL query below each answer, and the one-sentence summary of what the result shows.
- Not AI: the numbers in the result table, the parquet data, and every other page on TradeWeave (replications, research, country profiles, figures). Those are human-curated, cited to primary sources, and re-derivable from the SQL.
- Validation: the generated SQL is checked to be read-only and single-statement before running (
SELECT/WITHonly, 15-second timeout, 25-row cap). - Privacy: your question is sent to Anthropic to generate SQL. Do not paste confidential input.
- Training data used to build the workbench: none from language models. Every parquet table is a direct export of a primary source (CEPII BACI, CEPII Gravity, World Bank WDI, World Bank Pink Sheet, UNCTAD TRAINS NTM, WITS, ESCAP, FAOSTAT, IMF BoP, FRED, BLS, TRADHIST). No synthetic or AI-generated data.
Try an example
Requires ANTHROPIC_API_KEY in .env. The model generates SQL, which is validated (read-only, single-statement) and executed via DuckDB against the workbench Parquet tables. Nothing is written back.