Reader Search via Turbopuffer -- Motivation

Reader's current search is local-first, powered by a client-side SQLite database that indexes all documents locally. This architecture causes major issues: search freezes on large libraries, poor result quality, high memory/CPU usage, long indexing waits before users can search, and a 20-result limit. The team has already built performant Turbopuffer indexing for Reader MCP, so the infrastructure exists to transparently switch Reader search to server-side Turbopuffer FTS -- solving quality, performance, and scalability issues while opening the door to searching over more fields and returning better results.

Reader Search via Turbopuffer -- Shape (WIP)

This document is underway while the project is in Shape. It will be finalized when the stage completes.

Entered Shape: 2026-04-12 · Owner: @mati · Initiative: rw-projects

Goal

Determine feasibility and approach for switching Reader search from local SQLite to server-side Turbopuffer FTS, then define scope for build.

What's been done

• Notion spec written by Tristan: overview, benefits, future opportunities, technical notes
• Reader Search issues and feedback doc compiled (cataloging known problems)
• Turbopuffer document indexing already built for Reader MCP (Piotr's work) -- infrastructure exists
• Channel created, key people added (Mati, Piotr, Kris, Dan, Tristan)
• Piotr offered to help Mati with Turbopuffer internals

What's in progress

• Mati doing technical exploration and POC building
• Evaluating search quality: building test queries on large libraries, benchmarking old vs new results

What's not done yet

• POC demonstrating Turbopuffer search quality vs current SQLite search
• Offline fallback strategy (local RxDB filtering?)
• Rollout plan (opt-in beta setting vs full switchover)
• UX scope: which issues from the feedback doc to address in v1 vs defer

Open Questions (for next stage)

• How to handle offline search? Fall back to local filtering with user notice?
• Which UX issues from the feedback doc are addressable as part of the migration vs separate work?
• Performance characteristics of Turbopuffer FTS on very large libraries?
• Timeline for opt-in beta vs full rollout?