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

Deep technical conversation with the founder of Turbopuffer on building a vector/search database optimized for AI workloads, born out of Readwise's need for affordable semantic search (~$5k/month DB vs. ~$30k/month for vector search at scale). Covers the architecture bet on object storage + NVMe over traditional consensus layers, hybrid search design, and why models still need high-fidelity external retrieval systems. Highly substantive for engineers building RAG pipelines or evaluating vector DB infrastructure.