“Chat with your PDF” is now a familiar category. Upload a file, ask a question, and receive a polished response. It is convenient—and it quietly changes the task from reading a document to reading what a model says about it.
For orientation, that may be enough. For technical work, the path back to the claim matters. You need to know which page supports the answer, whether the wording came from the author, and what qualification sat just outside the retrieved excerpt.
Grounding is an interface decision
Grounding is not only a prompt or retrieval technique. It changes what the interface should do. An explanation should carry a source location. Selecting it should return you to the relevant page. A follow-up should remain attached to the passage that created the question.
The successful answer is not the one that ends the inquiry. It is the one that lets reading continue.
Local for the everyday questions
Many reading questions are small: define this term, restate this sentence, identify the contrast, or suggest what to notice next. On supported Macs, Pageflow can use Apple’s on-device language model for focused assistance without requiring an API key or sending the passage to a server.
Harder work—cross-document synthesis, larger context, or multi-step searches—may benefit from a cloud model. Pageflow keeps that choice visible. The useful distinction is not “AI or no AI,” but what leaves the device, which provider handles it, and what capability the user receives in return.
The reader remains responsible
No grounding system eliminates model error, extraction problems, or ambiguity in the source. Pageflow should make verification fast enough that checking becomes natural. That is a more honest ambition than making every answer sound certain.
AI belongs in Pageflow when it reduces the friction of understanding while preserving the reader’s relationship with the evidence.