Digital Green
Digital Green's Farmer.Chat answers farmers' questions in four countries
Digital Green, a nonprofit working with smallholder farmers, runs Farmer.Chat, a multilingual generative AI chatbot that answers crop and farming questions. Its team reports that close to 75% of questions were answered, with an average response time of 9.05 seconds.
Key figures
Copied exactly from the sources. Copy any figure with its citation.
Close to 75%
Share of user questions the platform answered1
71%
Average context precision (share of retrieved information that matches the question), over 1,000 queries tested1
Nearly 80%
Queries with high response faithfulness to the retrieved content (RAGAS evaluation)1
What it took
Cost, team size and timeline were not disclosed in the sources.
The problem
Smallholder farmers often cannot get timely, local advice on crops, pests and weather. Traditional extension services rely on agents visiting in person, which is hard to scale in remote areas. Earlier chatbots followed fixed scripts and handled few languages.
What they built
Farmer.Chat is a chatbot that answers farmers' questions in plain language. It runs on Telegram and WhatsApp and in a mobile app, accepts text, voice notes and images, and supports six languages (Swahili, Amharic, Hausa, Hindi, Odiya and Telugu, plus English). It is deployed in Kenya, India, Ethiopia and Nigeria. Users include farmers and agricultural extension agents.
How they did it
The system uses retrieval-augmented generation: it first looks up relevant passages in a local knowledge base of documents and videos, then asks a language model to write the answer from them. The paper says GPT-4 filters and re-ranks the retrieved material and GPT-3.5 handles other tasks. Weather forecasts and crop-disease diagnosis from outside services are added in some places. The first version launched on Telegram in Kenya in October 2023 and was improved over several rounds based on user feedback and qualitative studies.
Results
The team reports that the platform answered close to 75% of questions. Average response time was 9.05 seconds, mostly spent on re-ranking and generating the answer. In tests over more than 1,000 queries, average context precision was 71%, and nearly 80% of responses had high faithfulness to the retrieved content. 67% were rated highly relevant. The authors report high user satisfaction with response quality and relevance, particularly among women and low-literacy farmers, based on their user studies and interviews.
All reported figures
| Measure | Figure | Citation |
|---|---|---|
| Share of user questions the platform answered | Close to 75%1 | |
| Average response time | 9.05 seconds1 | |
| Average context precision (share of retrieved information that matches the question), over 1,000 queries tested | 71%1 | |
| Queries with high response faithfulness to the retrieved content (RAGAS evaluation) | Nearly 80%1 | |
| Responses rated highly relevant | 67%1 |
Lessons for your project
Ground answers in a curated local knowledge base. Farmer.Chat retrieves from region-specific documents and videos instead of relying on the model's general knowledge.
Go where users already are. Telegram and WhatsApp bots avoid asking low-literacy farmers to install anything, and voice notes and images widen access.
Measure unanswered questions and classify why. The team sorted misses into types, such as out-of-scope questions, to decide what content to add.
Test for bias before scaling. The team ran a gender-bias red-teaming exercise (deliberate probing for unfair answers) as part of the evaluation.
Caveats
- The figures come from a paper written by Digital Green staff and one Microsoft Research India author. They are not independently verified.
- Quality scores (context precision, faithfulness, relevance) come from automated LLM-based scoring (the RAGAS library), not from agronomists checking answers.
- The paper's headline user and query totals (over 15,000 farmers, over 300,000 queries) are not used as metrics because per-country counts in its table do not sum to the same total.
- The paper says GPT-4 is used for filtering and re-ranking and GPT-3.5 for other tasks. Models may have changed since October 2024.
- The industry tag is the nearest available fit for agricultural extension (farmer advice). The deployer is a nonprofit.
- The paper dates the first launch to October 2023 in Kenya. The paper has a stated date of 8 October 2024 (version 2).
Sources
- Farmer.Chat: Scaling AI-Powered Agricultural Services for Smallholder Farmers
arXiv (Digital Green and Microsoft Research India) · 2024-10-08 · Research paper · arxiv.org
Summarized in our own words from the sources above. Spotted an error? See corrections.