Development of a RAG-based LLM-Agent chatbot to support transformative adaptation planning for climate change

The accelerating impacts of climate change demand urgent and innovative responses to mitigate growing risks to both human and natural systems. Transformative adaptation planning offers a strategic framework to fundamentally reshape systems and governance structures, fostering long-term resilience and sustainability. In this context, Artificial Intelligence (AI) and Large Language Models (LLMs) present promising opportunities for enhancing knowledge discovery and distribution and supporting complex decision-making processes. This thesis presents the development of a chatbot designed to assist stakeholders engaged in transformative adaptation planning for climate change. The system leverages a Retrieval-Augmented Generation (RAG) architecture to generate informed and contextually relevant responses to user queries. A curated knowledge base, composed of textual documents related to the adaptation domain, is segmented into smaller chunks, embedded using a sentence embedding model, and stored in a vector database. Upon receiving a user query, the system retrieves semantically relevant chunks by computing their similarity to the query and incorporates them into a prompt for the LLM to generate a grounded response. The system reduces hallucinations by extracting information from the local knowledge base and enabling the LLM model to access up-to-date and domain-specific information without requiring fine-tuning. The chatbot is build upon an Agentic Workflow architecture, supporting the integration of additional functionalities such as chat memory and automated answer validation. The design of the chatbot aligns with the emerging field of LLM-Agents, which integrate planning, memory and external tools usage to perform complex tasks. The LLM operate as the brain of the system, controlling and orchestrating the flow of operations needed to complete the task. The resulting system is able to generate reliable answers to user queries, following a complex workflow that integrates at its core the retrieval of useful domain-specific informations from the knowledge base. Importantly, the presented approach uses relatively small language models locally and does not require model fine-tuning or retraining, significantly reducing its environmental impact by lowering energy consumption, CO2 emissions and computational resource requirements. The framework discussed in this thesis demonstrates the potential of LLM-driven tools to support knowledge-driven processes in transformative climate adaptation planning, highlighting their role in advancing AI-assisted Decision Support Systems (DSS) within the environmental domain.