Readwise turned to turbopuffer to ship AI-powered search features that wouldn't have been possible with any other tool from a cost and scalability perspective.
8TB+
vector and full-text search data
200k+
namespaces
1B+
documents
faster embeddings
Having our data in turbopuffer makes us feel ready for whatever comes next in AI-enabled software. It gives us the confidence to adapt quickly.
Tristan Homsi, Founder and CEO
Readwise is a reading platform that helps people collect, organize, and revisit books, articles, PDFs, and newsletters. The Readwise team needed a way for users to query across these entire libraries, not just single documents.
Other approaches fell short. Storing vectors in PostgreSQL wasn't scalable. Stuffing whole documents into context windows was too costly, limited by context size, and slow. Other vector databases were too expensive for a consumer subscription business.
turbopuffer gave Readwise a scalable foundation for retrieval. Even a user with 100,000 highlights can onboard immediately. This made advanced retrieval practical at consumer economics and allowed Readwise to focus on shipping highly requested features that were previously impossible to build, like allowing LLMs to reason with a user's book, article, and PDF highlights.
Readwise is special to the story of turbopuffer. Simon, turbopuffer co-founder, consulted with the Readwise team in 2022 to help them with a few infrastructure challenges. Simon worked with the team where they decided to not ship features because they were too costly to put in production on traditional search engines and vector databases. This was what inspired Simon was to build turbopuffer in 2023.
turbopuffer powers Readwise's new AI feature: Chat with Highlights. It enables readers to query the highlights they've saved from books, articles, and other sources:
Once the data is stored, turbopuffer retrieves it at scale. Readwise uses hybrid search to combine vector and full-text matching, ensuring queries return both semantic and exact keyword results quickly and accurately.
Looking ahead, Readwise is preparing to launch Chat with Your Entire Library. This feature expands retrieval from highlights to the full text of every book, article, and newsletter within a user's library, a scale increase of 10 to 100 times more data.
Readwise implemented turbopuffer in less than one month and experienced:
turbopuffer scales effortlessly, even when we throw a hundred thousand data points at it at once. We've never had issues with downtime or rate limits; it just works.
Tristan Homsi, Founder and CEO
Readwise migrated to turbopuffer's native embeddings in their
latest
web chat feature.
They embed queries on turbopuffer without re-embedding already indexed
documents. Their indexed columns use qwen3-embedding-8b with 4096 → 1536
MRL truncation. turbopuffer
automatically truncates returned embeddings to the dimensions of the vector
column, so they don't have to manage query embedding truncation on their own.
Switching to turbopuffer's native embeddings cut median embedding latency by 8x:
Readwise p50 embedding latency
previous │░░░░░░░░░░ 1.58s
tpuf │▓ ~0.2s
We will continue to update this log as Readwise's usage of turbopuffer evolves.
