Retrieval After RAG: Hybrid Search, Agents, and Database Design - Simon Hørup Eskildsen of Turbopuffer - Latent Space: The AI Engineer Podcast Recap
Podcast: Latent Space: The AI Engineer Podcast
Published: 2026-03-12
Duration: 1 hr 1 min
Guests: Simon Hørup Eskildsen
Summary
Simon Hørup Eskildsen discusses the evolution of TurboPuffer, focusing on the unique architecture that leverages object storage and NVMe SSDs for efficient database solutions, and the importance of search in AI-driven applications.
What Happened
Simon Hørup Eskildsen shares his journey from Shopify to founding TurboPuffer, motivated by challenges he faced with Elasticsearch. He discusses the importance of building a database that leverages modern storage architectures, highlighting the advantages of using object storage and NVMe SSDs. Eskildsen explains how TurboPuffer's architecture allows it to handle large data sets efficiently, with a focus on search and retrieval for AI applications.
The conversation delves into the importance of search in AI, particularly as it pertains to large language models and vector databases. Eskildsen outlines how TurboPuffer is designed to handle both full-text and vector searches, aiming to become the go-to search engine for unstructured data. He discusses how the company's architecture is optimized for read-heavy workloads, which are common in AI-driven applications.
Eskildsen elaborates on the conditions necessary for building a successful database company, emphasizing the need for new workloads and storage architectures. He highlights the importance of being able to implement every query plan over time, ensuring that a database remains relevant as technology evolves. He also discusses the challenges of maintaining a simple architecture while meeting the complex demands of AI applications.
The episode explores Eskildsen's experiences with early customers like ReadWise and Cursor, who helped shape TurboPuffer's development. He recounts how working with these companies illuminated the need for cost-effective, scalable search solutions that can support AI applications. Eskildsen shares insights on how TurboPuffer was able to reduce costs for its customers by optimizing its architecture.
Eskildsen discusses the role of venture capital in TurboPuffer's growth, sharing his experiences with investor Locky and the importance of transparency and trust in investor-founder relationships. He reflects on the challenges of fundraising and the critical decisions that led to TurboPuffer's current success.
The episode concludes with Eskildsen's vision for TurboPuffer's future, including plans to expand its feature set and continue evolving its architecture to meet the demands of AI applications. He emphasizes the importance of maintaining focus and listening to customer needs as TurboPuffer grows.
Key Insights
- TurboPuffer's architecture efficiently handles large datasets by utilizing modern storage options like object storage and NVMe SSDs, making it a strong candidate for AI applications where quick data retrieval is crucial.
- The need for databases to accommodate both full-text and vector searches is becoming critical as AI technologies advance, and TurboPuffer aims to be the go-to solution for managing unstructured data efficiently.
- Building a successful database company requires anticipating new workloads and evolving storage architectures, yet maintaining simplicity is key to meeting the complex demands of AI applications.
- Early partnerships with companies like ReadWise and Cursor helped TurboPuffer refine its architecture to offer cost-effective, scalable search solutions, reducing operational costs for its users.
Key Questions Answered
What is TurboPuffer's approach to database design?
TurboPuffer leverages modern storage architectures like object storage and NVMe SSDs to create a database optimized for AI-driven search and retrieval, focusing on both vector and full-text search solutions.
How did Simon Eskildsen transition from Shopify to founding TurboPuffer?
Simon Eskildsen left Shopify after nearly a decade, motivated by the challenges of scaling databases like Elasticsearch, leading him to develop TurboPuffer and address the need for efficient search solutions in AI.
Why did TurboPuffer choose Locky as an investor?
Simon Eskildsen valued Locky's authenticity and the ability to have open and honest conversations, which was crucial in building trust and effectively navigating the challenges of a startup.