Pablo Castro describes intrinsic knowledge as information embedded in a model's parameters and extrinsic knowledge as the outside context an agent needs for a task. He traces the move from simple retrieval to organization-wide grounding across documents, communications, structured data and public sources.
Pablo Castro explains why retrieval benefits from combining methods rather than relying on vector similarity alone. He demonstrates a layered knowledge-base approach that exposes simple configuration while retaining lower-level indexing control. Agentic retrieval adds reflection on whether the information need has been satisfied, with a latency-versus-quality tradeoff.
Pablo Castro describes learned knowledge as improvements derived from observing work and evaluating alternative agent configurations. His example establishes a baseline, tests candidates against an evaluation and applies a better configuration. Reported retrieval improvements and optimizer behavior are presented as the speaker's examples, not independent catalogue benchmarks.
Watch on YouTube
