HiParse-RAG: A High-Fidelity Document Parsing and Hybrid Retrieval Multi-model Fusion Framework for Complex Academic Question Answering

The study proposed the HiParse-RAG framework for the CCKS2025 complex academic question answering task. The framework improved accuracy through high-fidelity document parsing, hybrid retrieval, and multi-model fusion. PP-StructureV3 was employed to convert PDF papers into structured data, retaining text, tables, formulas, and figure descriptions. Fixed-length and semantic chunking strategies were combined for efficient text segmentation. In the retrieval stage, a hybrid method integrated sparse BM25 and dense vector search, and a Cross-Encoder model was used for reranking. Large language models generated answers, and rule-based post-processing ensured output consistency. Multi-model fusion and cross-validation corrected high-confidence errors across different configurations. The framework achieved an accuracy of 0.5644 in the official evaluation. The results show that high-fidelity parsing, retrieval-augmented generation, and model fusion are effective for enhancing performance in complex academic question answering.