RAG Agriculture Knowledge Assistant

A smart document retrieval system designed for farmers and agricultural professionals. This tool aims to help users manage and query information about their land, stored in tools like dashboards or uploaded documents (e.g., PDFs, CSVs). While the current version focuses on retrieving information from uploaded PDFs, the system is envisioned to integrate with dashboard data and CSV files in future versions. The retrieval-augmented generation system allows users to ask questions and receive accurate, referenced answers, making it easier to find critical information quickly and efficiently.

Overview

Key Features

• Document Upload/Indexing: Automatically indexes PDF files, cleans text using PyMuPDF, and generates vector representations for semantic searching.
• Natural Language Queries: Ask questions in natural, everyday language and receive answers that take context into consideration.
• Source Attribution: Each answer contains references to the source documents, making them traceable and verifiable.
• Vector Similarity Search: Based on ChromaDB for efficient semantic document searches by meaning, not keyword.
• Integration of LLM: Using Ollama for local inference with Qwen2:1.5B, preserving privacy and avoiding API charges.
• Document Management: See, list, and delete indexed documents via a clean API.

Technical Architecture

• Backend: FastAPI with async support.
• Vector Database: ChromaDB for persistent embedding storage.
• Embeddings: Default ChromaDB embedding function.
• LLM: Ollama (local inference) with Qwen2:1.5B.
• Document Processing: PyMuPDF (fitz) for text extraction, LangChain for recursive chunking.
• Containerization: Docker + Docker Compose managed via Makefile.

System Performance

• Fast Indexing: Processes complex PDFs quickly using high-performance C-based extraction (fitz).
• Sub-second Retrieval: Retrieves relevant context chunks in <100ms via vector search.
• Streaming Responses: Real-time token generation for a smooth user experience.
• Scalable Storage: Persistent vector database ensures knowledge remains available across restarts.

Installation & Setup

The project is fully containerized. Use the provided Makefile to manage the application:

# Build and start the entire stack (Ollama, Backend, Frontend)
make

# Stop all services
make down

# View live logs
make logs

API Endpoints

POST /upload # Upload and index PDF documents

GET /documents # List all indexed documents

DELETE /documents/{filename} # Remove document from knowledge base

POST /chat?query=… # Ask questions and get answers

POST /search?query=… # Vector search without LLM generation

Use Cases

• Agricultural Research: Query farming best practices, crop disease information, and irrigation guidelines
• Knowledge Management: Convert static document libraries into interactive Q&A systems
• Decision Support: Get quick answers from policy documents, regulations, and technical manuals
• Educational Tools: Create study assistants from textbooks and research papers

Technical Highlights

• Recursive Chunking: Uses a 1500-character window to preserve semantic meaning better than fixed-size blocks.
• Source-Grounded Responses: LLM is constrained to answer only from provided context, preventing hallucinations
• Containerized Orchestration: Seamless deployment of AI, Backend, and Frontend services using Docker.
• Local Data Privacy: Entire pipeline runs locally; no data ever leaves your machine.

Project Learnings

This project demonstrates expertise in:

• Information Retrieval: Vector embeddings, semantic search, and similarity metrics
• NLP Engineering: Text preprocessing, chunking strategies, and prompt engineering
• System Design: Building production-ready APIs with proper error handling and validation
• AI/ML Integration: Combining multiple AI components (embeddings, vector DBs, LLMs) into a cohesive system

Future Enhancements

• Multi-modal support (images, tables from PDFs)
• Conversational memory for follow-up questions
• Batch document processing
• Advanced filtering (by date, document type, custom metadata)

This project showcases the practical application of RAG architecture for domain-specific knowledge extraction, demonstrating how modern AI can make large document collections instantly queryable.