The full demo: 17 sources, 148 results, three-pass re-ranking, and Fetch More. With narration.
The question behind most enterprise search projects is not "can we build a search box?" It is "how long until someone can actually find the current version of anything?" When the answer involves copying every silo into a central index first, the honest estimate is months. This demo compresses the alternative into 49 seconds.
What happens in the video
A single query, data retention policy, goes out to seventeen sources at the same time: OneDrive, Outlook, Teams, SharePoint, YouTrack, GitHub, the open web, and more. Each source is queried live, in its own native API, under the searcher's own permissions. Twenty seconds later there are 148 results in one ranked list. Nothing was copied to get there, and nothing has to be kept in sync afterward.
Why the ranking is trustworthy
Federation has a hard problem hiding inside it: every source scores results its own way, and none of those scores are comparable. SWIRL solves it by re-ranking everything itself, in three passes. Lexical signals catch exact matches and phrases. Semantic embeddings catch results that say the right thing in different words. A cross-encoder then reads the query and each candidate together for the final ordering. The effect is that a Teams message, a SharePoint PDF, and a web page compete on what they say, not on which system happened to score them generously.
Pagination that goes back to the source
The end of the demo shows the part that surprises people: Fetch More. In a federated system, "page two" is not sitting in an index waiting. SWIRL remembers which sources still have results, re-queries exactly those, and merges the new page into the ranked list in place. In the video that is 31 additional documents, ranked and interleaved, nine seconds after the click.
What you did not see
No connector ETL project. No index rebuild. No staleness window between a document changing and search finding the new version. The corpus is wherever it already lives, which is also why deployment is measured in days, not quarters.
Read more in the documentation, or talk to us about running it on your own sources.