This paper presents a hybrid approach to automated answer generation in Russian-language chatbots, combining classical retrieval and neural network text generation methods. The system integrates a TF-IDF-based retrieval module with cosine similarity to identify the most relevant pre-existing answers in a knowledge base and a fine-tuned ruT5 transformer model to generate responses when no suitable template is found. A specialized corpus of question-answer pairs was prepared to adapt the ruT5 model to the domain-specific style and vocabulary. The experimental evaluation compared the performance of TF-IDF retrieval and the generative model using BLEU, ROUGE-L, and BERTScore metrics. The results demonstrate that the fine-tuned ruT5 model significantly outperforms the retrieval baseline, achieving improvements of 62.2% in BLEU, 3.7% in ROUGE-L, and 8.3% in BERTScore F1. The proposed architecture combines the high speed of retrieval with the flexibility of neural text generation, providing accurate and contextually relevant answers to user queries. The hybrid system can be effectively integrated into real-world customer support services, and its modular design enables further extension to multi-step dialogue processing and multimodal input handling.