Select Items: check two documents, Generate, and the citations are exactly those two. With narration.
There are moments when you do not want a retrieval algorithm making the call. Your insurance is up for renewal, the broker has questions, and the answer must come from the policy you actually hold, not from an insurer's marketing page that happens to rank well. This demo is about that moment.
The setup
A federated search for general liability insurance policy returns what real searches return: carrier product pages, industry news, case law from the web, and, further down the list, two documents that matter, the company's actual policy PDF in SharePoint and the broker's renewal email in Outlook.
The override
Toggling Select Items puts a checkbox on every result. Check the policy and the email, click Generate, and SWIRL builds the AI context from those two documents and nothing else. The answer comes back with the details that only the real policy contains: the carrier, the policy number, the limits, the effective dates. The citation list is exactly the two documents that were checked.
Why this is a guarantee, not a hope
Rerankers reduce the odds of the model reading the wrong thing; they cannot make the odds zero. Selection can, because it is structural: the retrieval set is your selection, so there is no step left where an off-target document can slip into the context window. When an answer will be forwarded, filed, or acted on, that certainty is worth two clicks.
When to use which
Most of the time, SWIRL's automatic path is the right one: it pre-selects its best matches, and canonical document election handles the "which version is current?" problem for you. Select Items is for the cases where you already know which documents govern, and the job is to make the AI read precisely those. The two compose: automatic retrieval to find, manual selection to decide.
The full walkthrough is in the RAG Guide. To see it on your own documents, talk to us.