How to Build Multimodal RAG with ColQwen2, Milvus, and Qwen3.5

This post details the implementation of a multimodal RAG pipeline using Milvus Lite, ColQwen2 for image-based vector encoding, and Qwen3.5 for generation. It demonstrates how to process PDF pages as images, encode them into multi-vector embeddings using ColQwen2, store these embeddings in Milvus Lite, and retrieve relevant page images based on a query. The retrieval logic uses a MaxSim aggregation strategy to score pages based on the highest-scoring patches for each query token. The retrieved page images are then sent to a multimodal LLM (Qwen3.5) for answer generation, showcasing a practical application for handling documents with tables, charts, and scanned content.