Abstract: In the context of digital transformation, project evaluation tasks face challenges such as rapidly growing data scales and significantly increased content complexity. Retrieval-augmented generation (RAG) introduces an external knowledge base for vector retrieval before generating model responses. However, to achieve high accuracy, the RAG process incorporates complex branches, multi-layer filtering, and rule matching, resulting in cumbersome workflows and high deployment barriers. To address these issues, this paper proposes a rule-based decision framework for multi-expert collaborative evaluation based on MAPPO (Multi-Agent Proximal Policy Optimization). This framework included modeling the project evaluation process and implementing a multi-dimensional reward mechanism to dynamically optimize rule selection and retrieval strategies. An attention-based decision layer was designed to enhance the model's focus on key rules and knowledge fragments. Experimental results show that the proposed method demonstrates stable convergence trends across rule sets of various scales. Visualization analyses reveal that in both early and later stages, different experts gradually develop complementary and differentiated choices, highlighting significant semantic division of labor and interpretability.