Retrieval After RAG: Hybrid Search, Agents, and Database Design — Simon Hørup Eskildsen of Turbopuffer

Turbopuffer's origin story and architecture are explored with Simon Eskildsen, delving into its evolution as a search engine for unstructured data. Eskildsen details Turbopuffer's unique approach to database architecture, leveraging NVMe SSDs and object storage, and its reliance on S3's consistency for consensus. He recounts leaving Shopify to consult for Readwise, where the high cost of embedding articles sparked the idea for a more cost-effective solution. The conversation highlights Turbopuffer's early customers, including Cursor and Notion, and the lengths they went to, such as buying dark fiber, to meet Notion's latency requirements. Eskildsen also shares insights into Turbopuffer's pricing strategy, team-building philosophy centered on "P99 engineers," and future plans, including expanding full-text search capabilities and scaling to handle massive datasets.