Inputs
Example prompts
- “Predict the most likely locations for a despondent 75-year-old last seen heading south from this point”
- “Run a prediction for a dementia patient, age 80, in this 2km radius area”
- “Generate a search priority map for this polygon area, the subject was last seen at 10pm”
- “Predict with heavy rain and strong wind conditions for this search area”
What you get back
A rich response (default) returns the following keys instructuredContent:
top_results— the top 50 H3 hexagons ranked by score (highest first). Each row carrieshex,score, and the cell centroid aslat/lngso the result can be chained directly intosearch_places,reverse_geocode,get_directions, etc. without a separate hex-to-coordinates step.scoreis an integer 0–30 — a relative confidence score across the surviving hexes in this prediction, NOT a probability or percentage.cells— every scored hex (the full grid). Each row carrieshexandscore. Suppress withinclude_hexes: falseif you don’t need the full surface in the structured response — the rendered hex layer inrenderis independent of this field.summary—hex_count(total hexes scored) andprocessing_time_ms.coverage—top_n(how many rows are intop_results),total_count(how many were scored),coverage_percent(score mass held by the returned rows), andtruncated(false when every scored hex was returned). Usecoverage_percentto gauge how concentrated the prediction is — a high value means the top hexes hold most of the likelihood; a low value means the search area is genuinely diffuse.render— a singleh3-cellslayer ready to drop into a map.style_hints.value_rangecarries[min, max]so the renderer can pick a colour scale. See render envelopes.
Response controls
LLM agents that don’t render maps and only want to chain into other tools should pass
compact: true. Renderer clients should leave the defaults.