Introducing Octen Image Search: Search for AI Enters the Multimodal Era

Humans drew before they wrote. Tens of thousands of years before the first written word, they were already recording the world on cave walls. Yann LeCun once estimated that by age four, a child has absorbed as much information through vision alone as all the text used to train the largest language models.

The web tells the same story. Text captures only a small part of the world, about 10% of the information online. The public web now holds more than 10 trillion seconds of video and more than 10 trillion images, and adds over a trillion new images every year. The look of a product, the streets of a city, a scientific diagram, a moment in history: all of it lives in images, frame by frame.

AI can now read images and draw them. It builds web pages, reports and slide decks. Yet the world still reaches it as text: URLs, titles and snippets.

Search is AI’s window onto the world. AI has entered the multimodal era. Search has to enter it too.

So we started building multimodal search for AI, search that lets models retrieve the visual information beyond text and see the world in full. Octen Image Search is our first multimodal search product. Give it a line of text or an image, and it returns real, relevant images from across the web, each with the page it came from.

From seeing the world to showing it

AI now speaks in more than text. It generates images, builds web pages and illustrates its answers. To show the real world, AI first has to see real images. When a model generates an image, those images are the reference it rebuilds the world from. When it answers a question, the image is part of the answer itself. That makes image search a foundational layer for AI applications.

Drawing the world as it really is

An image model’s world ends at the edge of its training data.

Ask it for a product or a character that appeared after its training cutoff, or for anything relatively obscure, and it has only a text description to work from. So it guesses, and the result often bears little resemblance to the real thing. Even for products, places and people that really exist, the details drift when the model works from memory alone.

Octen Image Search shows the model the world beyond its training data. In the grounded generation pipeline, search comes first. It finds the images the prompt calls for and hands them to the model as references. New products, new characters, architecture diagrams from papers: each comes out the way it really looks, and the scene stays true to the real world. Image and video generation both improve when they start from search.

Answers you can see

Most chatbot and agent answers are still plain text. Connect Image Search, and whenever AI needs an illustration or a visual reference, it searches, brings back relevant images with their source pages, and puts the images straight into its final output. The conversation becomes multimodal. Text explains, images show, and the answer becomes more complete, and more credible.

Search can start from an image, too. Upload a photo, and Image Search returns the original pages where it has appeared, plus other images of the same subject. Those pages often hold the story behind the image. With them, AI can identify what is in the photo, understand the world it captures, and answer questions about it.

Connecting visual information across languages and modalities

AI serves a global audience. The images it needs come from everywhere.

Octen Image Search understands natural-language questions in many languages and accepts an image as the query. Every result comes back structured, ready for an AI application to render directly in its front end.

The truest view of the world

AI’s view of the world is only as real as the images it finds.

In generation, the model rebuilds the scene from its reference images. The more accurate the references, the closer the result comes to reality. In conversation, the image reaches the user exactly as returned, and the user can tell at a glance whether it matches the question.

Image search also spans a vast range of scenarios and languages. That raises the bar for semantic understanding.

So we built an evaluation around how AI applications actually use image search. It covers 400 queries across 10 scenarios, from people, landmarks and products to logos, explanatory diagrams and news events, in 12 languages and a range of question styles. On the same queries and rubric, Gemini and Claude each scored every image that Octen Image Search, Google Images and Brave Search returned, on five dimensions: subject relevance, informativeness, image-text consistency, image quality and source authority. We then weighted the scores by scenario type and image rank to produce an overall score.

Octen Image Search scored 0.814 overall, the highest of the three.

Take the first image alone. In Octen Image Search’s results, it was highly relevant to the question 90.8% of the time, ahead of Google’s 85.0% and far above Brave’s 60.0%.

For AI, the first image is often all it needs. That means less filtering and fewer tokens spent.

Get started

Image Search is available to all users today, at just $2 per 1,000 calls, for any scenario that calls for image search.

Integration is a single POST request to https://api.octen.ai/image-search. Or type a question or upload an image in the Console and see what Octen Image Search brings back.

Try in Console · Read the docs · API reference