Climate Service Recipes: automatic multi-hazard climate information workflow generation using agentic Large Language Models (LLMs) and knowledge graphs

Climate Service Recipes (HACID-CSR) is an agentic system designed to assist providers of climate services in developing their advice for a wide range of clients. HACID-CSR guides providers by navigating the large and ever-increasing corpus of knowledge as well as an area without established standards and with limited access to scientific experts. It automatically generates detailed workflows (or “recipes”) by leveraging both a large language model’s internal reasoning and contextual knowledge from a domain knowledge graph for climate services (CS-DKG). The CS-DKG is an expert-curated ontology of climate service concepts with mapped relationships between climate variables, emission scenarios, indices, hazards, sectors, and key datasets (CORDEX, CMIP5, UKCP18), built as part of the Horizon Europe-funded HACID project (Hybrid Human Artificial Collective Intelligence in Open-Ended Decision Making).

The HACID-CSR architecture consists of a memory-enabled supervisor agent orchestrating multiple specialised agents. A planning agent first proposes an initial workflow outline, and a preliminary recipe agent uses only the LLM’s knowledge to draft answers to key workflow steps. The system then engages a knowledge graph retrieval sequence: a class selection agent identifies relevant classes in the CS-DKG, an instance selection agent finds specific instances (entries) highly relevant to the query within those classes following a two-stage selection process, i.e. semantic similarity based pre-selection and LLM-enabled refined selection, and a subgraph extraction agent retrieves the corresponding subgraph of related knowledge entities. Next, a recipe generation agent creates each step of the workflow by combining the LLM’s reasoning with the retrieved graph context using graph retrieval-augmented generation (GraphRAG). Finally, a recipe refinement agent compares the preliminary LLM-only solution with the knowledge-enhanced solution and refines the output, yielding a diverse and context-aware workflow.

By using this multi-agent approach, HACID-CSR increases the diversity of solutions and fills the knowledge gap between climate information and domain specific applications, helping experts to identify suitable methodologies and datasets. The resulting workflows are more traceable and transparent, improving user trust compared to answers from a general-purpose chatbot. We have also developed a bespoke automatic evaluation method to complement human expert validation of the generated recipes. We highlight the potential of the HACID-CSR approach for multi-hazard climate service design, and discuss remaining challenges and opportunities for further refinement of this agentic LLM-based system.