Abstract: Generative models like GPT-4, Claude, and PaLM have high levels of coherence and amazing reasoning capabilities, but they model knowledge "implicitly," making them vulnerable to breaking down in out-of-domain settings. To build a conversational assistant / chatbot that is aware of new, domain-specific knowledge, we will explain the necessity of Retrieval Augmented Generation (RAG) and the role of domain-adapted hybrid semantic search. We will also explain how RAG approaches also protect against the privacy and sensitivity implications of fine-tuning generative models, by applying permissions at the appropriate input layer. We will also overview how LLM Agents and a Tool framework build off of RAG to increase the space of solvable workflows and intents.
Learning objectives: Generative AI, LLMs, RAG, semantic search, LLM agents
Bio: Eddie joined Glean as a founding engineer. He leads a team of engineers and AI experts at Glean who work on challenging problems across Machine Learning, LLMs, NLU, domain adaptation, semantic search, and much more in order to build the world’s best work assistant. Previously, he worked as a research and software engineer for Google, contributing to a number of major projects, including natural language and semantic solutions. His current mission is to leverage AI and NLP technology to create the enterprise solution that speaks the language of each individual workplace.