VisRAG: Vision-based Retrieval-augmented Generation on Multi-modality Documents
This addresses the limitation of current RAG systems for handling multi-modality documents, offering a more effective solution for tasks involving visual and textual data, though it is incremental as it builds on existing RAG and VLM concepts.
The paper tackles the problem of text-based retrieval-augmented generation (RAG) systems being unable to utilize visual information like layout and images in multi-modality documents, introducing VisRAG, a vision-language model-based RAG pipeline that directly embeds documents as images for retrieval and generation, resulting in a 20-40% end-to-end performance gain over traditional text-based RAG.
Retrieval-augmented generation (RAG) is an effective technique that enables large language models (LLMs) to utilize external knowledge sources for generation. However, current RAG systems are solely based on text, rendering it impossible to utilize vision information like layout and images that play crucial roles in real-world multi-modality documents. In this paper, we introduce VisRAG, which tackles this issue by establishing a vision-language model (VLM)-based RAG pipeline. In this pipeline, instead of first parsing the document to obtain text, the document is directly embedded using a VLM as an image and then retrieved to enhance the generation of a VLM. Compared to traditional text-based RAG, VisRAG maximizes the retention and utilization of the data information in the original documents, eliminating the information loss introduced during the parsing process. We collect both open-source and synthetic data to train the retriever in VisRAG and explore a variety of generation methods. Experiments demonstrate that VisRAG outperforms traditional RAG in both the retrieval and generation stages, achieving a 20--40% end-to-end performance gain over traditional text-based RAG pipeline. Further analysis reveals that VisRAG is efficient in utilizing training data and demonstrates strong generalization capability, positioning it as a promising solution for RAG on multi-modality documents. Our code and data are available at https://github.com/openbmb/visrag.