Atlassian's Rovo semantic search has evolved from MiniLM to BGE-large to EmbeddingGemma-300m to handle real-world enterprise queries across Jira and Confluence. The system uses hybrid retrieval combining neural embeddings with structured signals like recency, project context, and engagement history. A collaboration with NVIDIA using the Llama-Nemotron-Embed-1B-V2 model fine-tuned on Jira-like data delivered 26–40% uplift in retrieval quality (Recall@60 and NDCG@1) in under one day. Upcoming investments include tenant-specific embedding fine-tuning, proactive assistive behavior, deeper work-graph modeling, and agentic AI integration via NVIDIA NemoClaw.
Table of contents
Why traditional search breaks down at workSemantic search in Rovo: finding answers, not just matching wordsInside “real‑world” semantic searchWhere we’re headed nextSort: