Accurate and efficient failure diagnosis in microservice systems is crucial; however, existing methods face two significant challenges: extracting semantic features from multimodal data and performing robust failure diagnosis with incomplete modalities. To address these challenges, we propose MultiCall, a novel framework for failure diagnosis that leverages Multimodal data analysis. MultiCall tackles semantic feature extraction by employing domain-specific bidirectional Mamba modules. It employs temporal feature modeling via masked event sequence training and contextual semantic reasoning augmented by domain knowledge (documentation/LLM-generated explanations), yielding instance-level representations weighted by abnormal scores. For incomplete multimodal data analysis, MultiCall utilizes a graph convolution network to model spatiotemporal dependencies, while a random modality masking strategy learns inter-modal relationships via projections for modality compensation. Extensive evaluations on three open-source platforms and a real-world microservice system demonstrate MultiCall’s superior performance. It achieves high accuracy in root cause localization, with A@1 (88.2%) and A@5 (95.6%), and excels in failure type identification, achieving an F1-score of 95.2%.