Visual Document Retrieval

Visual document retrieval searches documents such as forms, reports, and manuals by matching a query to evidence in page images, including text, tables, figures, and layout. Common approaches include dense single-vector embeddings, late-interaction multi-vector models that compare query and page patches, and OCR-based or sparse-text systems, often combined with reranking. Work on benchmarks such as ViDoRe and MMDocIR also explores layout-aware representations, token compression, and retrieval across multiple pages, with retrieved evidence often used in document question-answering systems. The central tension is preserving fine-grained evidence and covering all relevant pages while keeping indexes, latency, and inference costs manageable; performance can also vary with language, document type, and query complexity.