Video-based teacher online learning enables teachers to engage in reflective practice by watching others' classroom videos, providing peer feedback (PF) and reviewing others' work. However, the quality and reliability of PF often suffer due to variations in teaching proficiency among providers, which limits its usefulness for reviewers. To improve the quality of PF, hybrid intelligence is a promising approach that enhances human evaluation with AI-based techniques. Thus, we developed a hybrid intelligence feedback (HIF) system, where PF is categorized and summarized by large language models (LLM), and accompanied with AI multimodal data analysis, all in accordance with a knowledge structure. To investigate the effectiveness of the HIF, we conducted a study involving 58 pre-service mathematics teachers. After their initial feedback provision on a classroom video, they were divided into two groups. One group received HIF, while the other received traditional PF. Both groups revised their initial feedback based on the same video with PF or HIF report, and were assigned two tasks, namely in-depth reflection and extensive reflection. We analysed the reflective writings generated in in-depth reflection using the Structure of Observed Learning Outcomes taxonomy, and examined the diversity of teachers' attentions in extensive reflection using information entropy. Compared to traditional PF, our findings indicated that HIF (a) facilitated more comments added in feedback revision, (b) promoted multi-structural and extended abstract level thinking in in-depth reflection, (c) encouraged more diverse attentions in extensive reflection. These results demonstrate the effectiveness of HIF in enhancing PF to promote reviewers' reflection. This efficacy can be attributed to the utilization of LLM to identify common elements within PF, guided by the human knowledge-based framework, as well as the integration of data-driven evidence to complement PF.