Summarizing long-form texts like books presents significant challenges due to vast information, complex narratives, and long-range dependencies, often exceeding standard model capacities. This problem is particularly pronounced for Vietnamese, where dedicated resources and effective systems for book-length summarization remain underdeveloped. This paper introduces two contributions to address this gap: first, a new dataset specifically curated for Vietnamese book summarization, comprising diverse book content and corresponding reference summaries; second, a novel multi-stage hybrid pipeline designed explicitly for summarizing these long Vietnamese texts. Our pipeline employs a recursive approach, segmenting books into manageable chunks, iteratively summarizing them using a hybrid strategy that integrates extractive methods for salience identification and abstractive large language models for fluent generation, and finally reaggregating content while preserving narrative context. Pre- and post-processing steps ensure data quality and output refinement. We evaluated our system on the newly introduced Vietnamese book dataset using ROUGE metrics and human assessment. Experimental results demonstrate that our hybrid pipeline, particularly the LexRank and Qwen2.5-7B-Instruct combination, significantly outperforms purely extractive or abstractive approaches, validating the effectiveness of our proposed method and the utility of the new dataset for this challenging task.