Retrieval-Augmented Generation effectively mitigates hallucination issues in model generation by integrating the intrinsic knowledge of large language models with external knowledge bases. The performance of the retriever directly impacts the response quality of RAG systems in knowledge-intensive scenarios. However, existing retrieval algorithms and models predominantly designed for general domains, often exhibiting limited performance in specialized fields. To address this issue, this paper proposes a domain-specific knowledge retrieval framework named MMKRF, which accomplishes retrieval tasks in the field of material mechanics through three stages: recall, multi-path fusion, and re-ranking. Optimization strategies such as explicit knowledge enhancement, domain-specific vocabulary expansion, and domain-adaptive fine-tuning are employed to enhance the algorithms and models involved in the framework. Experimental results demonstrate that MMKRF exhibits excellent domain adaptability: on the material mechanics course knowledge retrieval dataset, MMKRF achieves a notable 5% improvement in accuracy (Acc) compared to baseline frameworks, while Recall@5 and MRR@5 increase by 1.3% and 3.4%, and Recall@10 and MRR@10 improve by 0.7% and 3.4%, respectively. These results underscore the framework's robust recall capabilities and its excellent top-ranking performance, enabling the provision of high-quality retrieval result sets for RAG systems.