Current mainstream SQL generation approaches remain insufficient in capturing the semantic information of structured data and handling complex query tasks. To address the challenges of hallucination and accuracy degradation in large language model (LLM)-based SQL generation, this paper proposes an enhanced SQL generation framework that integrates knowledge graphs with large language models. The proposed method introduces an SQL-KG-Verifier module, which synergizes the structured information of knowledge graphs with the generative capabilities of LLMs. By incorporating the verifier, the framework collaboratively refines SQL statements to ensure higher structural consistency, improved accuracy, and enhanced interpretability. Specifically, the verifier employs the entities and relational information retrieved from the knowledge graph as proxies to validate and revise the model outputs, effectively reducing generation errors. Experimental results demonstrate that on the BIRD and Spider datasets, the proposed method achieves execution accuracies of 52.71% and 77.26%, respectively—representing improvements of 19.93 and 21.27 percentage points over baseline models. Moreover, the approach exhibits superior adaptability and generation performance in complex and domain-specific query scenarios.

This study focuses on the crucial aspect of lossless compression for FAST pulsar search data. The deep generative model PixelCNN, stacking multiple masked convolutional layers, achieves neural network autoregressive modeling, making it one of the most excellent image density estimators. However, the local nature of convolutional networks causes PixelCNN to concentrate only on nearby information, neglecting important information at greater distances. Although deepening the network can broaden the receptive field, excessive depth can compromise model stability, leading to issues like gradient degradation. To address these challenges, this study combines causal attention modules with residual connections, proposing a residual causal attention module to enhance the PixelCNN model. This innovation not only resolves convergence problems arising from network deepening but also expands the receptive field. It facilitates the extraction of crucial image details while capturing the global structural information of the image, significantly enhancing the modeling capabilities for pulsar data. In the experiments, the model is trained and validated using the HTRU1 dataset. This study compares the average negative log-likelihood score with baseline models like the GMM, STM, and PixelCNN. The results demonstrate the superior performance of our model over other models. Finally, this study introduces the practical compression encoding process by combining the proposed model with arithmetic coding.

Retrieval-Augmented Generation (RAG) systems are increasingly adopted in government services, yet different administrations have varying customization needs and lack standardized methods to evaluate performance. In particular, general-purpose evaluation approaches fail to show how well a system meets domain-specific expectations. This paper presents CuBE (Customizable Bounds Evaluation), a tailored evaluation framework for RAG systems in public administration. CuBE integrates large language model (LLM) scoring, customizable evaluation dimensions, and a bounded scoring paradigm with baseline and upper-bound reference sets, enhancing fairness, consistency, and interpretability. We further introduce Lightweight Targeted Assessment (LTA) to support efficient customization. CuBE is validated on GSIA (Guizhou Provincial Government Service Center Intelligent Assistant) by using four state-of-the-art language models. The results show that CuBE produces robust, stable, and model-agnostic evaluations while reducing reliance on manual annotation and facilitating system optimization and rapid iteration. Moreover, CuBE informs parameter settings, enabling developers to design RAG systems that better meet customizer needs. This study establishes a replicable paradigm for trustworthy and efficient evaluation of RAG systems in complex government service scenarios.

The study focuses on the crucial aspect of lossless compression for FAST pulsar search data. The deep generative model PixelCNN, stacking multiple masked convolutional layers, achieves neural network autoregressive modeling, making it one of the most excellent image density estimators. However, the local nature of convolutional networks causes PixelCNN to concentrate only on nearby information, neglecting important information at greater distances. Although deepening the network can broaden the receptive field, excessive depth can compromise model stability, leading to issues like gradient degradation. To address these challenges, the study combines causal attention modules with residual connections, proposing the Causal Residual Attention Module to enhance the PixelCNN model. This innovation not only resolves convergence problems arising from network deepening but also widens the receptive field. It effectively utilizes global features, particularly capturing vertically correlated features prominently present in subgraphs of candidates. This significantly enhances its capability to model pulsar data.In the experiments, the model is trained and validated using the HTRU1 dataset. The study compares the average negative log-likelihood score with baseline models like GMM, STM, and PixelCNN. The results demonstrate the superior performance of the our model over other models. Finally, the study introduces the practical compression encoding process by combining the proposed model with arithmetic coding.