Collaborative filtering methods such as matrix factorization (MF) techniques are popularly used for building recommender systems because of their simplicity and effectiveness. However, matrix factorization techniques require a huge number of parameters as well as to include new users in the system, and the system needs to be optimized for the new user first. Neural network-based models, especially variational autoencoders (VAEs), are used for top-k recommendations to overcome the limitation of MF-based models. Mult-VAE and similar models such as RecVAE, H + Vamp are the state-of-the-art (SOTA) models for recommender systems currently. In this paper, we propose a method to consider the side information in the Mult-VAE model. We have used side information as the genres of the movies liked by the user. This side information is provided to the model by the means of a dense vector obtained from the term frequency of the side information. We decided to enhance Mult-VAE as it is simple and easy to accommodate the modifications to it compared to other SOTA models that use VAEs. Experiments are performed on the side information-based MovieLens 20 M dataset. The inclusion of side information contributed in improving the NDGC@100 metric compared to the base Mult-VAE model. This method can further be used for other networks as long as the base model follows encoder-decoder format as the VAEs. We show that the results of a VAE-based model can be improved by using the side information obtained by using the term frequency vector generated from the genre of movies.