Hybrid RAG multimodal knowledge service system architecture for underground coal gasification

Underground coal gasification (UCG) is pivotal for the clean and efficient in-situ utilization of deep and unminable coal seams. Its research demands high knowledge accuracy. However, general large language models (LLM) are limited in offering reliable knowledge services in UCG research and engineering due to insufficient specialized corpora, limited factual accuracy, and hallucination. This study therefore addresses the practical demand for multimodal knowledge services in underground coal gasification (UCG). We constructed a multi-source and multimodal dataset covering 8 categories of data throughout the full lifecycle of UCG, and established a domain-specific database through tokenization, standardization and quality evaluation. Precise representation of text modality was achieved through model fine-tuning and keyword enhancement. Textual conversion and semantic embedding were conducted for unified expression of numerical modality. Contrastive Language-Image Pre-training (CLIP) model was employed for feature extraction and cross-modal correlation of image modality, effectively eliminating the semantic gap between different modalities. We thus proposed a lightweight 3-layer UCG multimodal knowledge service framework featuring data representation, retrieval, storage and application based on the general Retrieval-Augmented Generation (RAG) architecture. A hybrid RAG process was designed by integrating sparse retrieval and dense vector retrieval. Ablation experiments show that the proposed hybrid RAG framework improved answer accuracy from 61.3% of the RAG-free baseline to 84.6%, increased factual consistency score from 3.1/5.0 to 4.2/5.0, and reduced hallucination rate from 38.7% to 16.1%. The proposed framework provides traceable knowledge services for UCG research.