Ask a question across a whole patient file
A file is thirty documents from eight sources over four years. This reads them all and answers a question across the set, naming the document and page behind each part of the answer.
- Finds a value by its number and a description by its meaning.
- Two documents agreeing are both cited, because that is corroboration.
- Recognition and retrieval both run on this device.
Across documents, disagreement is the finding
Within one report, two passages saying the same thing waste a slot. Across two documents they are the answer: a letter and a discharge summary agreeing on a date is corroboration, and the two disagreeing is exactly the thing worth surfacing. The rule that pushes out near-duplicates is therefore relaxed between documents rather than applied flat.
That was found the hard way. Undamped, a letter restating a date that a handbook gave differently scored as a near-duplicate, lost its place, and the answer named one date with complete confidence and no sign that anything else existed.
What it will not do
It does not interpret and it does not advise. It finds what the documents say and shows you where they say it, and every clinical judgement stays with a clinician.
It does no arithmetic. Values are quoted with their source rather than compared, averaged or trended, because a trend assembled from retrieved passages looks exactly like one somebody checked.
Where nothing relevant is found, the question never reaches the model and you are told so. A fluent answer built from the nearest available passage is indistinguishable from a real one, and on a patient file that is the single failure this could not survive.
Recognition and retrieval both run on this device. A patient file is the document people are least willing to upload and most often asked to send somewhere, and the usual resolution of those two facts is a free website and a shrug. Nothing here is uploaded, and nothing survives the tab being closed.
How many documents can I add?
A library of up to twenty files, with the page budget shown before anything is read.
Are the files uploaded?
No. Reading, recognition and searching all happen in this tab and nothing survives closing it.
Support and billing: support@gosmartpdf.com. Operated by Mrityunjay Kumar, India. Payments by Dodo Payments, our merchant of record.
Loading the editor…