In the Sefaria Library, users can use the Resource Panel to locate user-created Sheets from Voices on Sefaria that relate to a selected passage of text. When viewing Sheets related to a specific passage from the Library, users then have the option to sort Sheets by relevance — a feature that utilizes large language models (LLMs) to assess the content's relevance to the passage at hand. Alternatively, results can be sorted in two non-AI powered ways: by number of views or by publication date.

We took on this project to better support users in finding the content they're looking for. When learners look for Sheets through the Library's Resource Panel, they are looking for Sheets related to a particular passage of text. Therefore, we recognized that ranking results based on the actual content of the Sheets, rather than view count or publication date, would be a more effective way for users to locate the most meaningful content for their purposes.

Our experimental solution uses the following criteria to determine a Sheet's relevance:

1. Is that exact citation present in the Sheet?
2. How central is that citation to the overall content of the Sheet?
3.Besides the textual source, what other content did the Sheet creator include? The volume and nature of the other content is used to determine a creativity score.
4. How engaging and closely related to the passage at hand is the title?

Ultimately, this feature will help users more quickly and effectively find user-created content to aid in their exploration of original sources.

To learn more about Sefaria's approach to and uses of AI, visit Sefaria.org/ai