🌾 Krishi Sahayak

AI-Powered Farm Scheme Eligibility Assistant using RAG & Fine-Tuned LLM

Krishi Sahayak is an AI-powered assistant designed to help Indian farmers identify the government agricultural schemes they are eligible for. Instead of relying on static rule-based systems or generic chatbots, the project combines Retrieval-Augmented Generation (RAG) with a QLoRA fine-tuned Qwen2.5-3B language model to provide accurate, explainable, and context-aware recommendations.

The assistant retrieves the latest scheme information from a vector database and reasons over the farmer's profile before generating a response. When essential information is missing, the model intelligently asks follow-up questions instead of making incorrect assumptions.

🚀 Features

• 🌾 Government scheme eligibility prediction
• 🤖 Fine-tuned Qwen2.5-3B using QLoRA
• 📚 Retrieval-Augmented Generation (RAG)
• 🔍 ChromaDB vector database
• 🌐 Multilingual sentence embeddings
• 📄 PDF/knowledge retrieval support
• 💬 Intelligent clarification questions
• 📊 Evaluation pipeline
• 🎯 Gradio-based interactive chatbot
• ⚡ Google Colab compatible

🛠 Tech Stack

AI & Machine Learning

• Python
• Hugging Face Transformers
• PEFT (QLoRA)
• BitsAndBytes
• Sentence Transformers

RAG

• ChromaDB
• Cross Encoder Reranker
• Multilingual Embeddings

LLM

• Qwen2.5-3B
• LoRA Fine-Tuning

Frontend

• Gradio

Data

• JSON Knowledge Base
• Synthetic Instruction Dataset

📂 Project Structure

krishi-sahayak/
│
├── app/
│   └── app.py
│
├── rag/
│   ├── build_index.py
│   └── retriever.py
│
├── finetune/
│   ├── train_qlora.py
│   └── inference.py
│
├── data/
│   ├── schemes.json
│   └── generate_synthetic_data.py
│
├── eval/
│   ├── evaluate.py
│   └── test_profiles.json
│
├── requirements.txt
├── README.md
└── LICENSE

🏗 System Architecture

              Farmer Query
                    │
                    ▼
        ┌────────────────────┐
        │   User Profile      │
        └─────────┬───────────┘
                  │
                  ▼
        ┌────────────────────┐
        │   RAG Retriever     │
        │ (Chroma + Embedding)│
        └─────────┬───────────┘
                  │
          Relevant Documents
                  │
                  ▼
      ┌────────────────────────┐
      │ Fine-Tuned Qwen2.5-3B   │
      │