Conversational Search Redesign for a Note-Taking Startup Improves Recall by 27%

AI · 6 min read

Users wanted natural follow-up queries — for example, 'Show my note from last Tuesday about hiring' followed by 'Filter to candidates with design experience.' The original conversational layer treated each input as stateless, which led to failed queries and user frustration. The team implemented a session memory model that exposed context tokens in the UI and allowed users to pin or clear context with a single tap.

Designers iterated on micro-UX: a subtle 'context pill' bar above results, an undo for context changes, and a visual timeline of recent query anchors. Engineers optimized backend retrieval with context-aware embeddings to keep latency under 250ms. The team also added a short onboarding hint explaining conversational threads to make the affordance discoverable.

In lab tests with 60 participants performing multi-step search tasks, successful recall improved 27% and average task completion time fell by 18%. Participants reported the context pills felt empowering rather than intrusive when the pin and clear actions were obvious.