Recommender systems, particularly content-based (CB) recommendations, recommend items based on users’ past interests or preferences. Such an approach of recommendation triggers the issue of over-specialisation, which forms the phenomenon of homophily or filter bubbles. This phenomenon causes systems to recommend items that are too similar to those the users already know and that are not interesting. Serendipity refers to recommender systems’ ability to suggest relevant and novel items to users, often in unexpected or surprising ways. It’s about introducing users to items they might not have discovered on their own but are likely to find interesting or valuable. This research aims to address the issues of serendipity in CB recommendations by proposing and evaluating a few techniques, mainly based on the classic TF-IDF, the latent-Dirichlet allocation (LDA) topic modelling technique, and the knowledge graph (KG). During the experiment, the research used the MovieLens dataset and the Movie Plot Synopses with Tags (MPST) dataset, which contains plot synopsis data. Using the LDA technique, appropriate topics are generated from the plot synopsis text content and integrated into the knowledge graph along with other features. The proposed CB recommender system performs the recommendation in three steps: content analysis, profile learning, and filtering. For the analysis of results, apart from the standard precision and recall evaluation metrics, this research used the serendipity metrics, which consider popular items and items generated by a primitive recommender. A serendipity measure is used to gauge the serendipity level of items while maintaining relevance to the user. The results showed that using KGs can improve serendipitous recommendation performance compared to other models. The results indicate that representing users’ preferences and interests in the form of interconnected graphs results in the ability of the system to uncover serendipitous items.