KrishiSaathi
AI-Powered Multi-Agent Agricultural Intelligence
Process Flow Diagram
Query Processing Pipeline — End to End
How a farmer's question travels through the multi-agent system: from voice input to expert response in 3–5 seconds
Step 1 · Input
Farmer Input
Voice / Text in any of
12 Indian Languages
12 Indian Languages
Amazon Transcribe
Step 2 · STT
Speech-to-Text
Amazon Transcribe
10 Indian Languages
10 Indian Languages
deep-translator
Step 3 · Normalize
Normalize to English
Unicode script detect →
Internal processing language
Internal processing language
⚙️ Internal — for AI processing
Route to AI
Step 4 · Orchestrate
Supervisor Agent
Intent Classification · Entity Extraction
Memory Injection (12 memories) · Routing
Memory Injection (12 memories) · Routing
Crop Doctor
RAG: crop_diseases
farming_practices
farming_practices
Market Analyst
RAG: market_data
mandi_prices
mandi_prices
Scheme Advisor
RAG: government_schemes
schemes_database
schemes_database
Weather Prophet
API: OpenWeatherMap
RAG: farming_practices
RAG: farming_practices
Soil Expert
RAG: soil_data
farming_practices
farming_practices
Step 5 · Synthesis
Response Synthesis
Merge multi-agent outputs
Prioritize · Deduplicate
Prioritize · Deduplicate
Localize
Step 6 — Localization Layer
6.1
Translate to
Farmer's Language
Farmer's Language
deep-translator · chunk ≤4500
6.2
Text-to-Speech
Polly (Hi/En)
gTTS (8 langs)
gTTS (8 langs)
6.3
Store Context
RDS PostgreSQL
Mem0 Engine
Mem0 Engine
🌍 Output — for farmer delivery
Deliver
Step 7 · Output
Expert Response
Text + Voice in farmer's
language · Actionable advice
language · Actionable advice
3-5 sec
End-to-End Latency
92%
Intent Accuracy
12
Memories Injected
≥0.80
Keyword Skip Threshold
3
LLM Fallback Backends