Security: Load API keys from environment variable

Fix: Pass user query to LLM context in Fast Lane and Deep Dive stages

Add language matching: respond in same language as user query (Hindi/English/Hinglish)

Fix: Handle list type in context_priority_used extraction

Upgrade to Hybrid Architecture v3 (Fast Lane + Deep Dive)

Add 21 Groq API keys with fallback mechanism

Update Groq API keys with rotation logic

Update Groq API key

Update Groq API key

feat: migrate from Gemini to Groq API for LLM calls

Increase max_output_tokens to 4096 to prevent response truncation

Complete rewrite: direct API fetch, proper vegetation indices + cluster stress extraction

Fix: format specifier error in SAR section

Fix: extract Sentinel-2 vegetation_indices_summary and SAR crop_health

Improve prompt: force data-specific responses, add context logging

Fix: handle None values in SAR bands section

Fix: add defensive null checks in generate_response

Add comprehensive context: all indices, SAR, weather, persona-based responses

Switch to gemini-2.5-flash (2.0 quota exhausted)

Fix: use gemini-2.0-flash model

Add /chat/stream endpoint for Flutter compatibility

Add google-generativeai package

Restore original simple chatbot from initial commit

Simplify chatbot: single LLM call, remove complex reasoning

Fix: record_api_call once per query, not per key rotation

Pass skip_llm=True to Sentinel2, enforce 5 calls/minute limit

Add strict 5 calls/minute rate limit with smart fallback

Reduce API usage: default to simple reasoning (1 call), add backoff

Add SSE streaming endpoint /chat/stream for typewriter effect

Remove conversation history from LLM context to reduce tokens

Enhanced compact context: ALL 13 indices, clusters, prev_analysis, conv_history

Add compact prompts and context builder - reduces token usage by ~50%

Fix: add 400 to rotation triggers for invalid API keys

Fix: pass context to _generate_response for persona/zone/trend/weather

Add 5-key Gemini API fallback, conversation memory, historical trends, and zone recommendations

Lower confidence thresholds for satellite data fetch and reasoning

Fix maybeSingle not supported - use execute and check results

Fix user_profiles query - use address column instead of location

Full Developer Spec implementation: Priority-based Multi-Stage Reasoning

Add satellite band data injection from HF Space APIs (SAR, Sentinel-2)

Auto-fetch field data (coordinates, crop_type) from Supabase

Fix: Copy all Python files in Dockerfile

Implement multi-stage reasoning architecture (Claim→Validate→Contradict→Confirm)

Add auto model discovery to find working Gemini model

Try gemini-1.5-flash model

Switch to gemini-1.5-flash-8b for higher rate limits

Add retry logic for rate limits

Try gemini-2.0-flash-exp model

Fix: Use v1beta API and add better error logging

Fix: Use gemini-pro with v1 API endpoint