TLDR: Krishi Sathi is a new AI-powered chatbot designed to provide personalized, multi-turn agricultural advice to Indian farmers in English and Hindi, including speech support. It uses an instruction-tuned language model and Retrieval-Augmented Generation (RAG) on curated agricultural data to deliver accurate and context-aware responses, addressing limitations of previous systems and aiming to improve information accessibility in rural areas.
Indian agriculture, a sector vital to millions, often faces challenges in providing timely, accessible, and language-friendly advice to its farmers, especially in rural areas where literacy rates can be low. To bridge this critical information gap, a novel AI-powered agricultural chatbot named Krishi Sathi has been developed. This innovative system is designed to offer personalized, easy-to-understand answers to farmers’ queries through both text and speech, aiming to empower them with crucial agricultural knowledge.
Krishi Sathi stands out from traditional chatbots by adopting a structured, multi-turn conversation approach. Instead of merely responding to one-off questions, it engages in a dialogue to gradually gather necessary details from the farmer, ensuring a comprehensive understanding of the query before generating a response. This interactive method mimics a real-life conversation with an expert, fostering trust and making the system more approachable for users with limited digital skills.
How Krishi Sathi Works
At the core of Krishi Sathi’s intelligence is an Instruction-Fine-Tuned (IFT) language model, which has been specifically refined using extensive Indian agricultural knowledge. This knowledge is drawn from three carefully curated datasets, including information from reputable sources like the Indian Council of Agricultural Research (ICAR) and Vikaspedia.
Once the chatbot extracts the farmer’s intent and the context of their query, it employs a technique called Retrieval-Augmented Generation (RAG). This involves first fetching relevant information from a vast, curated agricultural database. Subsequently, the IFT model uses this retrieved information to craft a tailored, accurate, and reliable response. The system currently supports both English and Hindi languages, with integrated speech input (via Automatic Speech Recognition – ASR) and speech output (via Text-to-Speech – TTS) features, making it highly accessible for farmers who prefer speaking over typing and reading.
Performance and Impact
The development of Krishi Sathi has yielded promising results. The system achieved a Query Response Accuracy of 97.53%, demonstrating its ability to provide precise answers. Its contextual relevance and personalization stood at 91.35%, ensuring that responses are not only accurate but also highly relevant to the farmer’s specific situation. The query completion rate was also high at 97.53%, indicating that the chatbot successfully guides users to a resolution. Furthermore, the average response time remained under 6 seconds, ensuring timely support for users across both English and Hindi interactions.
This approach represents a significant advancement over previous agricultural support systems, such as the Kisan Call Centers, which faced limitations like inconsistent response quality and difficulties in handling query volumes. Unlike some earlier chatbots that were limited to specific crops or relied on static responses, Krishi Sathi’s RAG framework allows for dynamic access to a structured, domain-specific knowledge base, making it adaptable to various crops and regions.
Building the System: Data and Technology
The methodology behind Krishi Sathi involved several key stages. Data was meticulously collected from credible sources like ICAR, Vikaspedia, and specialized institutes such as the National Research Centre for Grapes (NRCG) and the Directorate of Onion and Garlic Research (DOGR). This raw data, initially around 20.4 million tokens, underwent a rigorous two-week curation process, resulting in a refined dataset of approximately 12 million tokens. This process involved automated preprocessing and human evaluation by agricultural experts to ensure high quality and domain relevance.
The curated data was then classified into distinct categories, defining 25 intents for grapes and 22 for onions, each with 2 to 5 slots to capture specific information like grape variety, climate, or fertilizer type. This structured approach enabled the generation of an instruction-based dataset for fine-tuning the language model.
For general-purpose functions within the system, such as routing queries or classifying crops and intents, the Param-1-2.9B-instruct model is utilized. For domain-specific agricultural queries, a fine-tuned version of the Param model is employed, leveraging in-context learning with few-shot prompts to generate precise and grounded responses. A dense vector retrieval module, built using the all-mpnet-base-v2 model and indexed with the Qdrant vector database, ensures that the model’s answers are factually aligned with domain-specific information.
Also Read:
- Developing Curriculum-Aligned Math Assessments Using Generative AI in Malaysia
- GanitBench: A New Benchmark Uncovers AI’s Multilingual Math Challenges
Future Directions
While Krishi Sathi shows strong potential, the developers envision several enhancements. Future iterations will include image analysis capabilities through platforms like i-SARATHI, allowing farmers to upload images of affected crops for visual diagnostic support. Integration with SAMBHAV will enable on-demand soil testing, leading to personalized nutrient management recommendations. The system also plans to integrate diverse agro-environmental parameters from IoT sensors and achieve seamless interoperability with existing government agricultural platforms like mKisan and KISAAN 2.0. Crucially, there are plans to expand multilingual support to include 22 official Indic languages, ensuring even broader accessibility for farmers across India.
This work demonstrates how combining intent-driven dialogue flows, instruction-tuned models, and retrieval-based generation can significantly improve the quality and accessibility of digital agricultural support in India. For more details, you can refer to the full research paper: Intent Aware Context Retrieval for Multi-Turn Agricultural Question Answering.
