The neural networks in Yandex Search have turned 10 years old.

Yandex introduced neural networks for image search exactly 10 years ago on December 5, 2014. Since then, they have been utilized in various fields, ranging from website ranking to text translation and object recognition in photos. Neural networks have evolved, becoming more accessible to a broader audience of users and companies. For instance, generative neural networks allow for the creation of images, texts, videos, and much more.

However, recognizing and searching for images—not only similar ones but also those closely related in meaning—was not the first feature that Yandex integrated neural networks into. As early as 2012, the company employed a simple neural network to predict traffic jams, and in 2013, it was used for speech recognition in the SpeechKit technology.

Then, in 2015, neural networks began to be applied in image search when processing text queries. Previously, the relevance of an image was determined by the surrounding text on the website, but the new model allowed for the evaluation of the image itself, placing it in the same semantic space as the text query.

In the search algorithm "Paleh," neural networks were first used for website ranking in 2016. The model developed by Yandex, similar to DSSM (Deep Semantic Similarity Model), helped assess the semantic relationship between webpage titles and user queries. A year later, in 2017, the "Korenov" update saw the neural network applied to page content, improving the quality of responses to unique queries.

In 2020, a heavy neural network, YATI (Yet Another Transformer with Improvements), was used for website ranking for the first time—an enhanced version of the "transformer," adapted for the "Search" runtime. This update significantly improved the quality of website ranking since the introduction of "MatrixNet" in 2009.

Yandex introduced neural networks into machine translation in 2017, allowing the "Translator" to consider context and translate phrases accordingly. The translation feature also became available in "Search": simply entering the query [translation перевод] yields an instant result.

In 2021, Yandex launched a full-fledged translation search: the idea is that if no suitable results are found in Russian, the system searches for them on English-language sites and offers translated options. That same year, a video translation feature was introduced in "Search" and "Browser," helping to expand access to useful information and overcome language barriers.

The primary goal of "Search" is to assist users in solving their tasks. To achieve this, Yandex provides not just a list of websites but also quick answers to questions, supplemented with links to sources. Previously, obtaining such concise answers required the YaLM language model, but starting in 2024, following the implementation of the new generation neural network YandexGPT, search has become more adept at analyzing complex questions and generating accurate responses.

Additionally, in 2024, Yandex introduced a multimodal VLM neural network into the search engine for the first time, combining expertise in working with text and visual models. Now, "Search" users can ask questions that combine text and images and receive detailed answers.