LocalGuide: On-Device Document Assistant

Problem

• People needed quick answers from their own documents, like a handbook or a set of policies, but reading through them by hand was slow.
• Cloud AI tools meant uploading private documents, which many people and teams cannot do.
• AI answers are easy to get but hard to trust when there is no source to check.
• Sensitive documents, and the questions asked about them, needed to stay private.

Solution

• Framed the on-device and offline constraints so documents and questions never leave the device.
• Designed cited answers as the core value, with the source passage shown behind every answer.
• Designed the three-pane workspace with a document library, a cited chat thread, and a source-passage panel.
• Set accessibility as an acceptance-level requirement, with every state shown in words and shape, never color alone.

Results

• Gave useful AI answers from a person's own documents without sending a file to the cloud.
• Kept documents and questions on the device, with nothing uploaded and no connection needed.
• Put the evidence up front by citing the source passage for every answer, so a reader can check it.
• Gave readers selectable answer styles to fit how they want to read.

What you can explore here

• The skeletal app below is a fully non-functional showcase of the interface. It is inert by design: there is no AI model, no upload, and no connection. Nothing you type is processed or leaves the device, the composer and the Add Document control are inert, and there is no file picker anywhere. The answers, citations, source passages, and documents shown are pre-baked sample data.
• Cited by design, private by design. Every answer carries the documents and passages it came from, and every state, on-device, offline, indexed, and the selected answer style, is shown in words and shape, never color alone.