TACO Demo is a research demo for asking questions about OpenStreetMap. Inspect the OverpassQL, explore the returned places, and see how in-context examples help translate a question.
Preview mode loads curated queries and dated OSM snapshots. It does not run an encoder, TACO selection or a language model. Connected queries run against the project’s own Overpass services.
Live mode embeds your question with the base and influence encoders, selects examples from the original pool with TACO, and sends the resulting ICL prompt to the configured OpenAI model, GPT-6 Luna by default. Selection keeps the original DeepSeek token budget; model API usage is reported separately. The ICL panel shows the returned similarity and influence scores, pool sizes, selector parameters and token counting method. Matching requests may reuse saved model or OSM responses; cache hits and historical model usage are labeled.
An independent research project using OpenStreetMap data. Not affiliated with the OpenStreetMap Foundation.
TACO Demo
Map workspace