How does Rovo decide which documents are the most relevant?

[AGASTYA SHARMA]

Hi everyone,

I've been reading about Rovo's search capabilities, and something I'm curious about is how it decides which Jira issues or Confluence pages to include when answering a question.

Is it mainly based on semantic similarity, keyword matching, or a mix of both?

And if there are dozens of relevant documents, how does Rovo rank them before passing them to the LLM?

I'm asking because understanding this would help me write better prompts and structure content more effectively.

Would appreciate any insights from the community or Atlassian team.

[Ankit Srivastava]

Hi @AGASTYA SHARMA 

Rovo Search retrieves relevant information from sources such as Jira, Confluence, and connected applications that the user has permission to access.

The Atlassian Engineering team has shared that Rovo uses multiple relevance signals to help rank search results, including content, metadata, recency, popularity, authority, contributor information, and user affinity. Rovo also processes content into smaller passages and ranks relevant information before generating responses.

However, Atlassian has not publicly documented the exact search algorithm, ranking formula, weighting of different signals, or the number of documents/passages considered when generating an answer. Details such as whether a specific result is selected based on keyword matching, semantic similarity, or a combination of approaches have not been fully disclosed.

Official references:

• What is Rovo?
  https://support.atlassian.com/rovo/docs/what-is-rovo/

• Search in Rovo
  https://support.atlassian.com/rovo/docs/search/

• Unraveling Rovo Search (Atlassian Engineering)
  https://www.atlassian.com/blog/atlassian-engineering/unraveling-rovo-search