feat: enrich Gemini with Florence facts and canonical damage labels

[madhavcodez]

What this does

Makes the Gemini-powered chat actually know things about Hurricane Florence and reliably do what users ask.

Three related changes:

1. Florence knowledge — added 8 grounded facts (category, landfall date/place, deaths, damage, rainfall, structures damaged, storm surge, dataset scope) to the system prompt so questions like "what category was it?" return accurate answers instead of a refusal or a hallucination.
2. Canonical damage labels — the filter tool now accepts both the canonical names Gemini learns from the prompt (no-damage, minor-damage, major-damage, destroyed, unknown) and the short aliases the frontend filter expects (none, minor, severe, destroyed, unknown). A small normalization map keeps both vocabularies consistent without breaking the frontend contract.
3. No more sterile empty replies — when Gemini returns a tool call with no text summary (an empirical flash-lite behavior), we synthesize a plain-English sentence from the action so the user always sees something meaningful like "Moving the map to that location." instead of "I processed your request but couldn't generate a text response."

Pinned to gemini-2.5-flash-lite because the full flash model intermittently returns 503 UNAVAILABLE under load — the lite model handles our tool-calling loop reliably.

How to verify

• "What category was Hurricane Florence?" → Category 4, weakened to Cat 1 at landfall
• "Take me to a destroyed building" → map flies to real coordinates, reply is a natural sentence
• "Show only severely damaged buildings" → filter applied
• "What's the weather in Paris?" → scope refusal

Addresses UTDisaster/frontend#48

[vercel]

The latest updates on your projects. Learn more about Vercel for GitHub.

Project	Deployment	Actions	Updated (UTC)
backend	Ready	Preview, Comment	Apr 14, 2026 3:57pm