The first phase of qualitative research is often hindered by the time-consuming process of transcribing and analyzing low-quality audio interviews. This paper introduces a novel human-in-the-loop pipeline that leverages Multimodal Large Language Models (MLLMs) and Retrieval-Augmented Generation (RAG) to overcome these. Our case study focuses on flood-related interviews conducted in North Bohemia, Czech Republic, in Czech, a language for which many AI models have not yet been optimized. Our proposed methodology employs MLLMs for both generating quantifiable, emotionally nuanced interview stimuli and for reconstructing interview transcripts from poor audio recordings. The initial AI-generated transcriptions are then refined and validated by a human, ensuring high accuracy and contextual integrity. Subsequently, a RAG system is utilized for efficient and transparent thematic analysis of the corrected transcripts. Our results demonstrate that this hybrid approach can achieve over 99% similarity between the AI-pipeline output and a manually corrected version, significantly reducing the time required for transcription and analysis by an estimated 75–80%. Furthermore, the RAG-based analysis enhances the rigor of qualitative research by improving efficiency, mitigating researcher bias, and ensuring auditability. This study showcases the potential of integrating advanced AI with human oversight to overcome data challenges and enhance the quality and efficiency of qualitative research, even in less-resourced languages.