The platform

BOOST-OI, a network supporting the agroecological transition in the Indian Ocean

BOOST-OI is a collaborative platform that identifies and connects stakeholders, projects, and documentary resources dedicated to the agroecological transition in the Southwest Indian Ocean and tropical zones. Led by CIRAD and its partners, it currently includes over 150 listed stakeholders, around 100 agricultural research and development projects, and nearly 220 technical and scientific documents, supplemented by a partner database of over 4,000 profiles (stakeholders, projects, intervention areas). The goal is to facilitate collaboration between countries in the Global South in addressing the challenges of sustainable development and adaptation to climate change. To further explore this documentary resource, BOOST-OI has deployed ChatBoosT, a conversational assistant based on artificial intelligence, allowing users to directly query the platform's resources using natural language.

  

 

 

 

 

 

 

Artificial intelligence to extract knowledge from BoosT-OI

ChatBoosT relies on the same technical principle as the Other CIRAD prototypes: RAG (Retrieval-Augmented Generation).

A language model alone sometimes invents answers when it doesn't know the exact answer to a pointed question—a phenomenon called hallucination. RAG corrects this problem by forcing the model to to rely on real documents before answering, rather than solely on one's training memory. In concrete terms, the process unfolds in three stages. First, indexing: each document in the BOOST-OI database (technical sheet, project report, publication) is broken down into fragments called chunks, then each chunk is transformed into a digital vector via vectorization (or embedding), which captures the meaning of the text rather than its exact words. These vectors are stored in a specialized database capable of quickly retrieving the passages closest to a question, even if worded differently from the source text. Next comes the semantic search: when a question arrives, it is itself vectorized, then compared to the set of chunks indexed by a module called a retriever, responsible for finding the most relevant passages — the information nuggets.

Finally, the augmented generation: the LLM (Language Model) receives the question along with the most relevant extracts found, and writes its answer based solely on this content — with the sources and citations to the support, so that each claim can be verified in the original document.

Easily understand how the RAG works in BoosT-OI by consulting the page Querying Knowledge in Agroecology  on the page of eKoal, our technical partner.