Linear's AI-powered similar issue detection helps customer experience teams consolidate backlog tickets with less manual effort
Using LLMs and vector embeddings at the scale of tens of millions of issues, Linear's Similar Issues feature surfaces overlapping requests during triage and at point of creation, reducing manual aggregation across engineering and support teams.
4 min read
Background
Large software teams accumulate duplicate issues, overlapping feature requests, and similar support tickets across engineering and customer experience systems. Manual deduplication is time-consuming and inconsistent, especially at scale.
What Was Implemented
- LLM-based semantic embeddings (cosine similarity) stored in PostgreSQL via pgvector on GCP
- Three integration points: issue creation modal, Triage inbox, Zendesk/Intercom support integrations
- Real-time surfacing of similar issues at point of creation and during triage
- Evolved into "Triage Intelligence" product on Business and Enterprise tiers
Results
Qualitative: reduced manual aggregation time for customer experience teams. No quantified metric (accuracy %, time savings) has been published by Linear. The feature operates at tens of millions of issues scale.
Lessons
- Vector embeddings enable semantic (not just lexical) similarity detection, catching duplicates phrased differently
- Integrating detection at the point of issue creation prevents duplication before it compounds
- Human review of AI suggestions preserves decision authority while reducing effort