Traditional Retrieval-Augmented Generation (RAG) systems often suffer from imprecise and overly broad context retrieval, which diminishes the quality and faithfulness of Large Language Model (LLM) answers, particularly with paraphrased queries. We introduce Entity Similarity RAG, a novel framework that enhances retrieval precision by leveraging entity-centric information from knowledge graphs. Our method includes two critical phases: an indexing phase constructs specialized entity-vector indexes and entity-relation key-value stores from source documents, and a query phase identifies entities in user questions, retrieves semantically similar entities via vector search, and fetches their explicit relational triples to form a highly relevant and structured context for LLM. Comprehensive experiments on the ComQA and GraphQuestions datasets, utilizing GPT-4o-mini and Gemini 2.0 Flash, demonstrate that Entity Similarity RAG significantly outperforms state-of-the-art baselines, including Naive RAG, GraphRAG variants, and LightRAG, across key RAGAS metrics such as context precision, answer correctness, and faithfulness. This entity-focused approach effectively addresses challenges posed by paraphrasing and reduces noise, leading to substantially improved reliability and accuracy in RAG systems.