AgriAdvice: Offline RAG chatbot to address the digital divide

App for smallholder farmers needing climate-adapted advice without internet. Scaled ambitious idea from a classmate ("AI on flip phone offline") to a technologically viable architecture. Uses RAG with a small language model to keep answers evidence-based. Optimized for extremely low-resource contexts. Focused on gaps like limited device access and unreliable internet connectivity.

How It Works

Designed in two phases for low-connectivity areas.

Setup (online, one-time, maybe intermittent upgrades):

• Extract text from 20-30 farming manuals using pypdf (text OCR) and pytesseract (image extraction).
• Embed 20-30 PDFs using Hugging Face models.
• Store in ChromaDB vector DB.

Offline use (laptops):

• small language model runs queries on embedded data. No internet needed.
• RAG setup: Model only uses retrieved docs to avoid hallucinations and false advice.
• All open-source, no vendor costs.

Tech Choices for Offline Architecture

• Small Language Model: Lightwaight enough for CPU on most laptops.
• RAG only: Sticks to input data to prevent hallucinations.
• Open source components only: Removes financial barriers and prevents vendor lock-in.
• English for MVP; will scale to multiple languages.
• Shared laptop target fits digital divide—avoids assuming personal phones.

Setup

# Install
git clone seohyeonlee2020/offline-rag-chatbot.git
pip install -r requirements.txt

# Embed PDFs (online)
python utils/text_data_preprocessing.py 

# Run offline on localhost
streamlit run agriadvice_main.py  

Next Steps

• Usage documentation
• Multilingual support
• Group easily serchable information into a mass SMS service to reach users who do not have access to computers.