PyCon Hong Kong 2025

Large language models face fundamental limitations, including outdated training data, context window constraints, and a tendency to hallucinate. While vector similarity search is a popular solution through Retrieval-Augmented Generation (RAG), simple cosine similarity often fails to capture the nuanced relevance needed for high-quality LLM responses.

This talk explores how search capabilities can enable your LLMs to provide accurate a. We examine how sophisticated retrieval methods serve as external memory, enabling LLMs to access vast knowledge bases beyond their training cutoff. Through hybrid search combining dense and sparse retrieval, semantic reranking, and query reformulation, we demonstrate how search becomes the critical bridge between knowledge and real-world information needs.