Gemini-GraphQA: Integrating Language Models and Graph Encoders for Executable Graph Reasoning

Graph-structured data presents challenges for natural language question answering due to its non-Euclidean topology and task-specific requirements. To solve this, we propose Gemini-GraphQA, a new graph question answering framework that combines a large language model (Gemini) with graph neural networks and retrieval-augmented generation strategies. Unlike traditional models that use shallow feature mapping or isolated code synthesis, Gemini-GraphQA uses a graph encoder to capture structural semantics, a graph solver network to translate natural language into executable graph code, and a retrieval module to add external knowledge to the reasoning process. An execution correctness loss is added to ensure the generated code is both syntactically and functionally correct, allowing the framework to outperform existing graph-based QA systems and pretrained code generation models. This design improves the model's ability to reason across various graph-related tasks and enables its deployment in fields requiring structured data understanding.