Visual RAG: when document images speak in responses

The problem: incomplete responses without visual context

Traditional RAG systems extract only text from documents. But technical manuals, product guides, and catalogs contain diagrams, screenshots, and visual tables that are essential for correct answers.

Concrete example: a software manual has screenshots of configuration screens. When a user asks "how to configure a cash register?", the textual answer is useful but seeing the actual screen is much clearer.

The solution: Visual RAG

Visual RAG extends the classic RAG pipeline with automatic image extraction:

Smart extraction

Not all images in a document are useful. The extractor applies automatic filters:

• Minimum size: Images smaller than 50x50px are ignored (icons, spacers)
• Solid color: White backgrounds or separators are discarded
• Deduplication: MD5 hash prevents repeated logos on each page
• Configurable limit: Maximum 50 images per document (adjustable)

Proximity association

Each image is associated with text chunks on the same page (PDF) or position (DOCX). This means when RAG retrieves a relevant chunk, images from that section automatically accompany the response.

The Image Manager allows manual adjustment: reassign images to other chunks, upload additional images, or remove irrelevant ones.

Display in chat and widget

Images appear inline in the response, below the LLM text:

• 1 image: Direct display with lightbox on click
• 2+ images: Carousel with navigation (arrows, dots, counter)
• Embeddable widget: Same behavior inside Shadow DOM

Zero latency impact

All heavy lifting happens in background during document processing. At runtime, only image URLs are passed — zero impact on response time.

Conclusion

Visual RAG transforms technical documents from "walls of text" into rich visual responses. Users see exactly the screen, diagram, or table they need, alongside the textual explanation.