KrishiBot: an NLP approach to enhance farming with low resource language

Abstract

The economy of Bangladesh has expanded significantly in recent years; however, farmers continue to lack sufficient training and timely assistance, which negatively impacts productivity. Given that many farmers reside in remote rural regions, accessing quick and accurate agricultural support remains challenging. This research addresses these issues by developing an advanced chatbot that functions as a virtual agricultural assistant, leveraging Naive Retrieval-Augmented Generation (Naive RAG), GraphRAG, and state-of-the-art Large Language Models (LLMs)—GPT 4.1Mini, GPT-4o, GPT-3.5 Turbo, LLaMA 3.3 70B, and Claude 3.5 Haiku—selected for their superior capabilities in human language comprehension, generation, and interpretation. These systems retrieve precise, domain-specific information from our fine-tuned databases using RAG and GraphRAG methods, delivering targeted solutions tailored to specific agricultural scenarios. The research employs LangChain as the orchestration framework, Pinecone as the vector database, Neo4j as the graph database, and multilingual-e5-large as the embedding model. Furthermore, prompt engineering bridges interactions between LLMs and Bengali-speaking farmers, providing targeted guidance on crop selection, weather conditions, and fertilizer usage for enhanced practical support. Experimental evaluations demonstrate GraphRAG’s superior performance compared to Naive RAG. Specifically, GraphRAG achieved 91.58% accuracy on a QA test set of 500 agricultural questions, outperforming Naive RAG’s 78.6%, and scored 84.85% in semantic similarity versus Naive RAG’s 80.66%. These results confirm the advantage of leveraging graph-based knowledge retrieval to provide more accurate, contextually relevant responses.