Application and Optimization of DeepSeek-Based Safe AI Framework in the National Power Grid Multi-Level Knowledge Base for Authority-Task Management

The security-oriented AI framework, built upon DeepSeek and multi-tiered knowledge storage, offers a solution for enterprises to deploy localized proprietary knowledge bases. Unlike cloud-based models, DeepSeek enables local model construction and training, integrating multi-level databases and stringent data management mechanisms to significantly reduce the risk of sensitive information leakage. This feature is particularly critical for high-security enterprises such as State Grid, allowing them to leverage AI for operational efficiency while ensuring data protection. However, the current multi-tiered knowledge base structure, which categorizes knowledge by tasks and access privileges, suffers from low retrieval efficiency due to the vast data volume. This inefficiency hampers timely access to optimal results, diminishing model practicality and potentially leading to misleading outputs or risks. To address this, we propose a hierarchical recommendation algorithm for external knowledge bases in large language models. By dynamically partitioning the knowledge base and leveraging user role-based permissions and historical dialogue records, the algorithm precisely targets relevant knowledge, enhancing retrieval efficiency and accuracy.