LightOn said it built “Visual RAG” into its AI platform on Feb 26, 2025, calling it a world first

Written by Claude Sonnet 5.5, an AI model, from public records. Each figure in it was checked automatically against the source's text or our database before it was published.

LightOn, a generative AI company that said it was founded in Paris in 2016, announced on Feb 26, 2025 that it had integrated “Visual RAG” into its platform. It described itself as the first European generative AI company listed on Euronext Growth, the Paris exchange’s market for smaller companies. The company called the step a world first: users could work with and interact with documents that mix text, images, charts and diagrams, with an AI that reads those sources together. The release is in French, and the AMF, France’s financial markets regulator, lists it as inside information, its category for announcements a listed company must publish promptly.1^,2^,3^,4^,5

The release describes Visual RAG as a new approach to multimodal retrieval-augmented generation (RAG), a way of having an AI model draw on a collection of documents when it answers. It said existing solutions were limited to understanding isolated images, while Visual RAG ran complex searches across text and visual elements, and that users could process millions of mixed documents.6^,7^,8

Igor Carron, LightOn’s co-founder and chief executive, said Paradigm, the company’s platform, was the first generative AI solution able to process and analyse images at such a scale, retrieving and processing millions of them within a mixed document base. He added that users could now talk with photos or infographics as they had with text since generative AI began.9^,10^,11^,12^,13

The company named reading financial reports, patents and technical manuals as uses, and said Visual RAG integrated directly into organisations’ own infrastructure, which it said kept their data confidential. It said the feature was available from that day for integration into information systems. LightOn tied the launch to its research and open-source work, naming ModernBERT, PyLate and MonoQwen.14^,15^,16^,17

The AMF lists the release twice, and we hold the text of both copies: 3 pages each in French (4,274 and 4,271 characters), with pictures and layout left out and no document the release refers to. In the text we hold, the company names no customer and states no price for Visual RAG.18^,19