How Rovo Deep Research works - Inside Atlassian How Rovo Deep Research works

This post details the introduction of 'Deep Research,' a novel feature within Rovo Chat that enhances information retrieval and report generation for complex inquiries. It elaborates on the enhanced RAG architecture, including query-dependent retrieval and content-grounded answer generation. Key technical contributions include the transition from single-path to multiple-path information retrieval by decomposing tasks into sub-tasks, and implementing structured report generation through outlining, parallel section development, and integration. The post also highlights the use of Atlassian's Teamwork Graph for foundational data, the iterative reflection and refinement of results, the generation of reports with in-line citations and exportability to Confluence, and the specific reasoning models employed (Llama 3.2 8B, GPT-4o, GPT-4.1, Claude 3.7 Sonnet). It also outlines the evaluation framework for measuring response quality and the challenges addressed, such as limited depth of insight and insufficient source attribution.