vidore

Another multi-vector, visual retrieval model based on Qwen2. The size of the model being only 2B parameters makes it a great choice for many RAG scenarios.

Based on SmolVLM, ColSmolVLM is a ColPali-style retriever model, which is becoming more popular in the retrieval community. Instead of extracting the text from documents, the models embed the whole document as a picture, therefore not only preserving the text but also figures and tables.