Rovo, the semantic search system powering Jira and Confluence, has improved document retrieval accuracy by 26.7%, enabling users to find relevant work artifacts faster through nuanced understanding of domain language and context.
- Recall@60 improved from 75.1% to 95.1% with domain-specific fine-tuning.
- Rovo’s semantic search understands work context beyond keywords.
- NVIDIA-powered model tuning completed in less than 24 hours.
What happened
Atlassian’s Rovo semantic search platform, integrated with Jira and Confluence, has achieved a 26.7% increase in relevant document retrieval performance. This was measured by Recall@60 metrics, showing an improvement from 75.1% to 95.1%, meaning the correct document appears within the top 60 results far more frequently.
This enhancement was realized by fine-tuning the Llama-Nemotron-Embed-1B-V2 model on a publicly available Jira-like dataset using NVIDIA’s fine-tuning recipe and AI software stack. The process, leveraging tools such as NVIDIA NeMo Data Designer and NeMo Automodel, took less than a day to complete and delivered between 26% and 40% uplift in retrieval quality, translating into more precise and context-aware search capabilities.
Why it matters
Traditional keyword-based enterprise search systems inadequately understand complex workplace language, resulting in lost productivity due to irrelevant search results and duplicated efforts. Rovo’s semantic search is designed to interpret the meaning behind queries, including the use of acronyms, project codenames, and partial information commonly found in work environments.
By tailoring search results to the context of real work scenarios and adapting models per organization using anonymized signals, Rovo significantly reduces friction in finding critical documentation, Jira issues, and decision histories. This capability empowers millions of monthly active users to work more efficiently, make informed decisions, and maintain project momentum without time wasted on sifting through poorly matched content.
What to watch next
Rovo is actively investing in expanding its ability to model the interconnections of work artifacts ranging from high-level initiatives to operational postmortems. Future enhancements aim to make search interactions more conversational and proactive, providing summaries, comparisons, and risk highlights directly within user workflows.
Additionally, further personalization through deeper per-tenant embedding adaptation will enable Rovo to better align with the unique language and workflows of each organization. Continued collaboration with partners like NVIDIA promises scalability gains in both semantic search quality and performance, reinforcing Rovo’s position as a cornerstone technology in Atlassian’s AI-powered productivity ecosystem.