A Gujarati AI WhatsApp chatbot for dairy farmers

The problem: advice that doesn't reach the farmer

A large Gujarat dairy cooperative wanted to give its member farmers quick, reliable answers: what to feed a cow after calving, when to call a vet, what their last milk payment was, when the next collection is. The information existed — in advisory documents, extension-officer knowledge and the dairy's own systems — but reaching thousands of farmers in villages, mostly in Gujarati and mostly on basic Android phones, was the hard part.

WhatsApp was the obvious channel. Farmers already use it every day, it works on low-end phones and patchy networks, and it supports voice notes and images as well as text.

How the chatbot works

1. A farmer sends a message on WhatsApp — Gujarati script, Romanised Gujarati, Hindi or English.
2. The WhatsApp Business API delivers it by webhook to our Node.js service.
3. The service decides what kind of question it is: general advisory, or something about the farmer's own account.
4. Advisory questions go to the RAG pipeline: the most relevant passages from the cooperative's approved content are retrieved and given to the language model, which writes an answer in Gujarati based only on that content.
5. Account questions (payments, collection data) are answered from the dairy's backend APIs after matching the farmer's registered phone number.
6. Anything the bot is unsure about, or anything that sounds like an animal-health emergency, is flagged with advice to contact a veterinarian or the local society.

Why RAG and not just ChatGPT

A general chatbot will confidently answer questions about cattle medicine from whatever it learned on the internet. For a cooperative, that is a liability. Retrieval-augmented generation keeps the model grounded: it can only answer from content the cooperative has approved, and it says so when the content doesn't cover a question. The advisory team can update the knowledge base without retraining anything.

Making Gujarati work properly

• Farmers mix scripts. Many type Gujarati words in English letters, so the bot handles Romanised Gujarati as well as Gujarati script.
• Answers are short and practical — a few lines, not an essay — because they are read on a small screen.
• Units and terms match what farmers use locally (litres, fat percentage, SNF), not textbook vocabulary.
• Template messages for reminders and payment alerts are written and approved in Gujarati.

Deployment inside the client's infrastructure

Large cooperatives often run their own data centres and Windows servers. The bot's API runs on Node.js hosted under IIS with iisnode on the client's main server, close to the backend systems it queries, while template messages are sent through their WhatsApp provider account. Keeping farmer data on the cooperative's own infrastructure was a requirement, not an afterthought.

What we would tell any business planning a WhatsApp AI bot

1. Start from the questions people actually ask — collect a few hundred real ones before writing any prompts.
2. Ground the model in your own content and systems; never let it improvise on safety-critical topics.
3. Design for the language people type, not the language in your documents.
4. Plan the handoff to a human before launch.
5. Measure: unanswered questions are the best list of what to add to the knowledge base next.

Planning an AI chatbot on WhatsApp in Gujarati, Hindi or English? SICI Technologies is an Official Meta Business Tech Provider — send us a short brief.