Topic analysis

How we index images for RAG

Kapa, a company building AI assistants for technical documentation, developed a cost-effective method to integrate millions of images (like screenshots, diagrams, and schematics) into their RAG pipeline. Instead of processing images per query with expensive multimodal models, they use a cheap vision model to generate detailed text descriptions at indexing time, storing these captions as retrievable text chunks—reducing per-query overhead by 1-6% compared to text-only systems, improving answer quality statistically significantly, and avoiding payload limits and recurring vision costs.