Stanford CME295 Transformers & LLMs | Autumn 2025 | Lecture 7 - Agentic LLMs

The lecture explores practical techniques for Large Language Models (LLMs) to interact with external systems, focusing on Retrieval Augmented Generation (RAG), tool calling, and agents. RAG is presented as a method to augment prompts with relevant information from external knowledge bases to overcome the limitations of LLMs' knowledge cutoff dates and context length constraints. The discussion covers chunking strategies, embedding models (like SentenceBERT), and retrieval methods, including semantic similarity search and BM25. Tool calling enables LLMs to complete tasks by accessing external resources through structured data and function APIs, exemplified by finding a teddy bear using location data. The lecture also introduces agents, autonomous systems that pursue goals through iterative processes, highlighting the ReAct framework and agent-to-agent communication protocols, while also addressing safety concerns like data exfiltration.