This study introduces a novel Large Language Model (LLM) driven framework for automated species grouping and diet matrix generation in Ecopath with Ecosim (EwE) ecosystem models, addressing a critical bottleneck in model development. The framework (i) retrieves a marine species list from an area; (ii) uses LLMs to classify them into functional groups; and (iii) synthesises trophic interactions from diverse data sources including global biodiversity databases, species interaction repositories, and unstructured user-provided text. We evaluate the framework across four large Australian marine regions to assess both consistency and ecological accuracy of the resulting functional groups and diet proportions. The framework demonstrates high reproducibility in species grouping decisions (>99.7% consistency) and diet matrix construction, with 51-59% of predator-prey interactions showing consistent diet proportions across multiple runs. Validation against expert-derived matrices for the Great Australian Bight ecosystem reveals strong ecological alignment and accuracy, with 92.6% of taxonomic assignments being at least partially correct (>75% fully correct), and correctly identifying 85% of trophic interactions, while estimating diet proportions within 0.2 of expert values for 80% of interactions. These findings demonstrate the framework\'s potential to generate reproducible, ecologically meaningful components for ecosystem model development while significantly reducing development time.