We're planning a live virtual event later this year and would love to hear from you. Are you using powerful AI technologies that you think everyone should be using? Here's your chance to show it to the world!
AI is too often seen as a business of the rich, by the rich, and for the rich. Let's take a look at Digital Green's Farmer.Chat, a generative AI bot designed to give small farmers in developing countries access to critical agricultural information. Developing countries have often implemented technological solutions that would never have occurred to engineers in rich countries. They solve real problems rather than appealing to venture capitalists’ “let’s start another Facebook” fantasy. Farmer.Chat is one such solution.
Learn faster. Take a deeper dive. Look further.
Farmer.Chat helps agricultural extension agents (EAs) and farmers get answers to their questions about agriculture. Deployed in India, Ethiopia, Nigeria, and Kenya. Although originally designed for EA, farmers are increasingly using them themselves. They are already accustomed to using social media to ask questions online. The clear goal was to provide fast, efficient online access to better, more reliable agricultural information.
AI applications for farmers and EA face many constraints. One of the biggest constraints is location. Agriculture is hyperlocal. Two farms may be a mile apart, but if one farm is on a hill and the other is in a valley, the soil, drainage, and weather conditions may be completely different. What works for your neighbors, with different microclimates, pests, crops, etc., may not work for you.
Data exists to answer local questions on topics such as fertilization and pest management, but it is spread across many databases with many owners, including governments, NGOs, and corporations, as well as local knowledge about effective methods. Farmer.Chat uses all of these sources to answer your questions. But doing so requires respecting the rights of farmers and database owners. Farmers have the right to privacy. They may not want to share information about their farm or tell others about the problems they are experiencing. Businesses may want to limit what data is exposed and how it is exposed. Digital Green solves this problem with FarmStack, a secure, open source protocol for opt-in data sharing. All connections use end-to-end encryption. All data sources, including farmers and government agencies, choose what data to share and how to share it. They may decide to share certain kinds of data and not other data, or they may impose limits on the use of the data (for example, limiting it to certain geographic areas). The granular selection sounds impressive, but by respecting data providers and users, Farmer.Chat has been able to build a trustworthy ecosystem for data sharing. As a result, that ecosystem leads to a successful farm.
FarmStack also allows for confidential feedback. Has the data provider's data been used successfully? Did the farmer provide local knowledge that would benefit others? Or was there a problem with the information? Data is always a two-way street. It's important not only to leverage data, but also to improve it.
The biggest challenge for Digital Green and Farmer.Chat is translation. Farmer.Chat currently supports six languages (English, Hindi, Telugu, Amharic, Swahili, and Hausa), and Digital Green is working to add more languages. To serve EA and farmers well, Farmer.Chat must support multiple modes, including voice, text, and video, and be accessible in the farmer's native language. Useful information is available in many languages, but finding it through voice chat and answering questions in the farmer's language can be very difficult. Farmer.Chat uses Google Translate, Azure, Whisper, and Bhashini (an Indian company that provides text-to-speech and other services for Indian languages), but there are still gaps. Even within a language, the same word can have different meanings to different people. Many farmers measure their yield in bags of rice. What is a “bag of rice”? It might be 10 kilograms to one farmer, or 5 kilograms to someone selling it to another buyer. Keeping the extension agent in the loop is an important area. EA can recognize issues with local usage, local slang, technical farming terminology, etc. and resolve them by asking questions and interpreting answers appropriately. EA also helps with trust. Farmers are understandably wary of taking advice from AI when it comes to changing practices that have been used for generations. An EA that knows the farmers and their history and can place the AI’s answers into local context is much more trustworthy.
To manage hallucinations and other incorrect output issues, Digital Green uses Search Augmented Generation (RAG). RAG is conceptually simple. It retrieves relevant documents and constructs a prompt that instructs the model to construct a response from those documents, but in practice it is more complex. As anyone who has ever searched knows, your search will likely return thousands of results. Including all results in a RAG query is impossible for most language models and impractical for the few language models that allow large context windows. Therefore, you need to score your search results for relevance. Be sure to select the most relevant document. You should then organize your document so that it contains only the relevant parts. Keep in mind that for Digital Green, this issue is multilingual and multimodal. Related documents can be displayed in any language or mode you use.
It is important to carefully test all steps in this pipeline, including the translation software, text-to-speech software, relevance scores, document cleaning, and the language model itself. Could another model do a better job? Guardrails must be installed at every step to prevent erroneous results. Results must pass human review. Digital Green tests with “Golden QA,” a highly regarded set of questions and answers. When asked the “golden question”, can the application consistently produce results as good as the “golden answer”? Testing like this must be done on an ongoing basis. Digital Green also manually reviews 15% of our usage logs to ensure that our results are consistently high quality. On O'Reilly's podcast, Andrew Ng recently noted that the evaluation phase of product development often doesn't get the attention it deserves. In part, this is because writing AI software is so easy. Who wants to spend months testing an application that took a week to write? But that's exactly what you need to succeed.
Farmer.Chat is designed to be gender inclusive and climate smart. With 60% of smallholder farmers around the world being women, it is important that applications welcome women and not assume that all farmers are men. Pronouns are important. The same goes for role models. Farmers presenting skills and answering questions in video clips should include men and women.
Climate-smart means making recommendations as climate-sensitive as possible. Climate change is a big problem for farmers. This is especially true in countries like India, which may be affected by rising temperatures and changing rainfall patterns. Recommendations must anticipate current weather patterns and how they may change. Climate-specific recommendations tend to cost less. Farmer.Chat, for example, isn't afraid to recommend commercial fertilizers, but emphasizes local solutions. Almost every farm has an unlimited supply of compost. It costs less than fertilizer and helps manage agricultural waste.
Farming can be very tradition-bound. “We do this because our grandparents and their parents did it.” New agricultural technologies developed by faceless scientists in city offices mean little. It's much more likely to be adopted if you hear about it being used successfully by a farmer you know and respect. To help farmers adopt new practices, Digital Green uses video collected from local farmers to prioritize the work of its peers whenever possible. They strive to keep farmers in touch with one another, celebrating their success in helping farmers adopt new ideas.
Lastly, both Farmer.Chat and FarmStack are open source. Software licensing may not directly impact farmers, but it is important for building a healthy ecosystem around projects that aim to do good. There are too many applications that aim to monopolize your attention, subject you to unwanted surveillance, or devalue political debate. Open source projects to help people: We need more.
In Farmer.Chat's most recent phase of history, Digital Green has supported more than 6.3 million farmers, increasing their incomes by up to 24% and crop yields by up to 17%. Farmer.Chat is the next step in this process. And we wonder whether the problems facing small farms in developed countries are no different from those in developing countries. Climate, insects and crop diseases have no respect for economics or politics. Farmer.Chat helps small farmers succeed in developing countries. The same services are needed in the so-called “first world”.