Function Calling for LLMs: RAG without a Vector Database

In this talk, Jim Dowling explores extending Retrieval Augmented Generation (RAG) with Function Calling to access structured/tabular data without relying on a vector database. The presentation covers how to enrich tables with metadata and the expressivity of queries that can perform well. It examines function calling in the context of queries to the Hopsworks feature store, which supports extensive metadata and statistics for columns and tables (feature groups) to improve function calling performance. The talk discusses both cloud-hosted LLMs (like GPT-4) and private LLMs, such as Hermes-2 (a fine-tuned mistral 7b LLM), demonstrating practical code examples and architectural patterns.