The paper presents the development of an automated information retrieval system of media content on user's request in natural language. It describes the current problems of search in the context of rapid growth of media content volume and diversity, and also reveals the limitations of traditional search methods based on the exact match of keywords and metadata. Special attention is paid to the analysis of natural language processing (NLP) methods, such as tokenization, lemmatization, vectorization and semantic proximity computation, as well as the use of neural network models based on the “transformer” architecture. A comparative analysis of modern open language models, including Qwen3, Vikhr, Saiga and YandexGPTLite 5, in terms of their applicability for multilingual and multimodal text retrieval and generation tasks is carried out. The paper proposes solutions for integrating modern NLP-methods and neural network algorithms into the server side of the search system, which allows to improve the relevance, accuracy and convenience of media content search for unstructured and imprecise user queries. The presented approaches provide a balance between search quality, performance and versatility of the system, and open up prospects for further development of intelligent search services.

Qwen3, Vikhr, Saiga, YandexGPTLite 5, NLP, relational attention module, pre-trained model, Transformer Seq2Seq, text to SQL conversion, Spider dataset.