While building Bodhi AI, integrating retrieval-augmented generation (RAG) posed several challenges, especially around ensuring retrieved context was both relevant and up-to-date. Early versions sometimes pulled outdated or irrelevant documents, leading to incorrect or nonsensical answers, and scaling the system to handle large volumes of real-time queries introduced latency. To overcome these issues, I implemented semantic caching, improved data preprocessing, and adopted efficient vector search algorithms, which significantly boosted both accuracy and speed. Tight integration between the retrieval and generation components required careful prompt engineering and output validation, but ultimately resulted in a robust, scalable solution that delivers real-time, reliable answers for users.