A Domain-Specific Multimodal Embedding-Based Integrated Retrieval System for Cybercrime Investigation

This study proposes a domain-specific multimodal embedding model and an integrated retrieval system for cybercrime investigation. To address the limited domain adaptability of general-purpose embedding models, we fine-tune Jina Embeddings v4 via LoRA using synthetic triplets generated through an LLM-based pipeline. The proposed system indexes structured and unstructured evidence in a unified Qdrant vector database and orchestrates retrieval workflows using LangGraph. For image evidence, we propose an Image-Text Fusion strategy that combines original visual embeddings with VLM-generated descriptions and OCR text. Under optimized hyperparameters, the proposed Fusion strategy achieves a Hit Rate@1 of 0.9407 and MRR of 0.9641, improving upon the baseline Jina v4 Single Vector strategy by 48.1%p and 42.2%p, respectively. Furthermore, the proposed Single Vector strategy achieves an MRR of 0.9509, offering a resource-efficient alternative with an 8x smaller index size and over 6x faster retrieval speed.