I indexed 936 Lex Fridman episodes into a RAG that cites its sources (github.com)

A new AI tool, OmniPod, has indexed 936 episodes of the Lex Fridman podcast, enabling users to query detailed, source-cited answers from the content. This system significantly reduces the problem of "hallucinations" commonly found in chatbots by ensuring every response is verified against exact transcript excerpts. Users can ask specific questions, such as what influential figures like Andrej Karpathy say about neural networks, and receive accurate, grounded responses instead of fabricated quotes. OmniPod utilizes a sophisticated retrieval-augmented generation (RAG) pipeline that distinguishes between factual, synthetic, and generative queries. It employs a specialized intent routing mechanism to optimize responses based on query type. With an embedding dimension of 384 and a rapid filtering system using Qdrant for vector searches, the application's performance is both fast and efficient—all without requiring a YouTube API key. This innovative approach not only enhances the reliability of AI-generated information but also sets a precedent for more accountable and precise interactions in the growing AI/ML landscape.