ColQwen2

ColQwen2 is a visual retriever in the ColPali family, published in June 2024 by contributors from ILLUIN Technology and Hugging Face. Instead of relying on OCR text extraction, it encodes each document page as an image with a Qwen2-VL backbone, capturing layout, tables, charts, and figures into multi vector embeddings scored with late interaction similarity.

The approach gives retrieval systems a fuller picture of visually rich documents such as PDFs, slide decks, and scanned reports. ColQwen2 was added to Hugging Face Transformers in 2025 with a dedicated ForRetrieval class and processor, making it straightforward to build image first RAG pipelines.