MSDA-Text: Template-Guided Long-Form Text Generation with Multi-Source Data Augmentation

Large Language Models (LLMs) have demonstrated remarkable capabilities in text generation, yet their outputs often depend heavily on pre-training data and lack the factual depth required for domain-specific long-form writing, such as industrial reports or biographical summaries. To address this limitation, we propose MSDA-Text (Template-Guided Long-Form Text Generation with Multi-Source Data Augmentation), a framework designed to produce accurate and comprehensive long-form texts aligned with user intent. Building upon existing long-text generation architectures such as Storm, MSDA-Text introduces two key enhancements: (1) a template-guided outline generation process that incorporates user-provided reference materials into multi-perspective LLM discussions, and (2) multi-source data augmentation that integrates both Internet-based and local real-time data through Retrieval-Augmented Generation (RAG) and Text-to-SQL techniques. The framework employs the Model Context Protocol (MCP) to unify template parsing across heterogeneous file types and features a long-text writing agent that autonomously retrieves and synthesizes content for each outline section. Experimental results demonstrate that MSDA-Text generates long-form documents that are more structured, user-aligned, and factually grounded than existing LLM-based methods.