Speaker
Description
The application of large language models (LLMs) in forestry is still experimental to large extent, specifically in highly specialized subdomains such as the European castaniculture. This scientific initiative presents CastaneaGPT, a domain-specific digital assistant which is designed to support sweet chestnut (Castanea sativa) management in Italy and beyond. CastaneaGPT is conceived as both a working prototype and a methodological case study, and aims to illustrate how a professional AI system can be developed within a forestry context.
The system follows a modular architecture with a clear separation between backend “reasoning” and frontend presentation. The backend is based on a Retrieval-Augmented Generation (RAG) pipeline that combines OpenAI LLM models with LlamaIndex for document ingestion, indexing, and retrieval. The integrated text corpus includes curated scientific literature, grey literature, and web-based materials, processed through structured PDF and URL ingestion workflows with metadata extraction and citation control. Upgrade measures such as enhanced retrieval strategies (e.g., MMR, hybrid BM25), chunk-size optimization, conversation memory, and a forestry-related glossary are implemented to consolidate the factual grounding of the tool. FastAPI serves as the orchestration layer, enabling containerized deployment (Docker) and flexible frontend integration (e.g., Glide). The user interface is planned to integrate three main features: dashboard, map (PostGIS-based), and chat components.
Beyond the technical realization, the project is driven by two principal scientific interests: (1) the methodological challenges in building a reliable forestry-specific RAG corpus, and (2) the empirical testing and evaluation of different interaction modes between user and AI assistant (e.g. for professional castaniculturalists; seekers of decision support when reactivating abbandoned chestnut plantations; playful users such as children) in order to evaluate the influence of these interaction modes on the comprehension and the trust of the users.
CastaneaGPT is conceived as a digital platform to connect the castanicultural community, support decision-making, and strengthen knowledge dissemination, while offering a reproducible blueprint for similar domain-specific AI tools in forestry.
| Keywords | Digital_Forestry; CastaneaGPT; LLMs; Decision_Support |
|---|
