Why Linear Moved Duplicate Detection to the Integration Edge
TRIGGER
Duplicate detection during issue creation or triage was helpful but came too late in the workflow—issues had already been created and entered the system before duplicates were surfaced, meaning support teams still had to context-switch between tools and manually aggregate related customer messages.
APPROACH
Linear extended their similar issues feature progressively closer to the source: first in issue creation, then in triage inbox, and finally directly in support integrations (Intercom) where customer emails arrive. Support teams see similar existing issues and their status directly next to incoming emails without leaving their support tool. The same vector embedding infrastructure powers all three touchpoints. Result: customer experience team consolidated support issues with less time spent manually aggregating messages.
PATTERN
“Every step further from the creation moment—integration to triage to backlog search—means someone already wasted time creating the duplicate. Detection in backlog search is cleanup; detection at the integration edge is interception. Optimize for interception.”
✓ WORKS WHEN
- Issues originate from multiple external sources (support emails, chat, form submissions) that can host integration UI
- High volume of incoming reports where even small percentage of duplicates creates significant overhead
- Support team is separate from engineering and lacks deep knowledge of existing backlog
- Integration platforms support embedding custom UI components or sidebar widgets
- Latency requirements for embedding lookup are achievable at edge (<500ms to not block user flow)
✗ FAILS WHEN
- Most issues are created directly by internal team members who know the backlog well
- Integration platforms don't support custom UI extensions or real-time data fetching
- Issue volume is low enough that manual duplicate checking is feasible (<50 issues/week)
- Duplicates primarily occur within a single session (same person creating duplicates) rather than across sources
- Support workflow requires immediate response before similarity results could be computed