Ask Predict Charts
Ask natural-language questions against the datasets behind Predict Charts. Coverage is deepest for Kalshi; cross-venue questions work where comparable data is available.
Ask questions in plain English against the underlying trade data.
How this works
Your question goes to a FastAPI server, which asks the configured LLM to generate read-only SQL, validates that SQL against a table allowlist and a set of syntax and I/O restrictions, runs it against DuckDB views over Parquet and CSV data, and returns the SQL, the rows, and an interpretation.
The dataset stays on the server. What is sent to the LLM provider is: your question, the schema description, and — so the model can interpret results rather than just emit SQL — up to the first 10 rows returned. “Deeper insights” sends up to 20 rows, plus a similar sample from each supporting query it runs. If retrieval-augmented context is enabled, your question is additionally sent to OpenAI’s embeddings endpoint, regardless of which provider generates the SQL.
Ask for a metric, population, and date window in the same sentence, for example "sports vs non-sports fees in February 2026." Use follow-ups for interpretation after the first query succeeds; the SQL and evidence panels are there to help sanity-check what the model actually used.