Three questions to the SWIRL Assistant, each answered from a live federated search, each cited. With narration.
Enterprise chatbots usually work from a copy: documents are ingested into a vector store, and the bot answers from whatever snapshot the last sync captured. The SWIRL Assistant is built the other way around. It has no copied corpus at all. Every question triggers a live federated search, run as the person asking, and the answer is synthesized from what that search returns, with citations.
Start with the revealing question
The demo opens with the question every assistant should be able to answer honestly: "What sources can you search for me?" The reply is not a canned capabilities blurb. It is the actual provider list for this deployment, grouped: Microsoft 365, Box, iManage, databases, the open web. If a new source is connected tomorrow, tomorrow's answer includes it.
Then the real work
"What are the limits of our general liability policy?" The Assistant searches, finds the policy PDF in SharePoint, reads it, and answers with the actual figures, per-occurrence and aggregate, cited to the document.
"What insurance does the Dakota Partners lease require us to carry?" This one lives deep inside a lease agreement. The answer covers the required coverage types, the carrier rating requirement, and the tenant's pro rata share, each traceable back to the lease itself.
What the architecture buys you
Because retrieval happens at question time, there is no staleness window: the answer is as current as the system of record. Because the search runs under the asker's identity, permissions are enforced by the sources themselves rather than re-implemented in a second security model. And because every answer carries citations, "trust me" is never part of the experience; the source is one click away.
Details are in the AI Search Assistant Guide. To put it in front of your own sources, talk to us.