Doximity Ask's PDF upload feature lets clinicians question uploaded documents rather than search every page manually. The useful output is not merely a fluent summary: it is an answer that preserves dates, missing information and the distinction between what the document says and what the model infers.
Doximity's 3 July 2026 feature announcement describes support for clinical documents, research papers and guidelines, with references to relevant pages and sections when applicable. This review checks that published description as of 6 September 2026. It does not report a live upload test or assume that every scanned page, table or attachment will be interpreted successfully.
Ask a question the document can answer
Start with a bounded request. "What changed between the referral and follow-up?" is more useful than "Tell me everything important" when your immediate task is to reconcile the two records.
State whether you want extraction, comparison or interpretation. Extraction asks what is present. Comparison identifies differences between source passages. Interpretation adds a judgement about what those differences may mean. Blending all three without labels makes it harder to check where the answer came from.
Before using identifiable information, establish that the product and account are approved for the intended clinical workflow under your organisation's arrangements. For training or evaluation, use synthetic documents. The ability to upload a PDF is not, by itself, permission to upload any document you possess.
Example one: a referral with conflicting dates
Consider an original fictional referral dated 2 September 2026. Its main paragraph says the symptom began in July, while an attached earlier note says it was first discussed in June. The referral also mentions an investigation but contains no result.
An appropriate test prompt would be: "Extract the stated symptom timeline, identify any conflicting dates and list investigations whose results are not included. Refer to the relevant source locations and do not resolve discrepancies by guessing."
The expected review task is clear. Check whether the output retains both dates, whether it labels the discrepancy and whether it treats the missing result as unavailable. It should not invent a reconciled onset date or infer a normal result from silence.
This is a constructed evaluation example, not an observed Doximity response. Actual results should be published only after the document is uploaded and the output reviewed against the source.
Example two: a laboratory report with an unhelpful summary
For a second synthetic exercise, create a report with samples collected on different dates, a corrected result and one pending item. No clinical interpretation or dose calculation is needed to test whether the document is read faithfully.
Ask for the collection date, reporting date, status and value as printed for each requested item. Keep these as separate fields. A table that places a result beside the wrong sample date can look orderly while being unusable.
Then ask which entries should not be treated as final. Check the answer against the correction notice and pending label. The critical distinction is not whether the tool can repeat a number, but whether it preserves the context that gives the number meaning.
Do not silently combine results from different encounters. If the source is unclear, the answer should retain that uncertainty rather than construct a tidy trend that the document does not support.
Example three: a research paper with several outcomes
For a research exercise, use a permitted paper or an original synthetic study summary. Identify the prespecified primary outcome, the result reported for it and any secondary findings that receive more attention in the discussion.
An appropriate prompt would request those items separately and ask the tool to distinguish reported results from the authors' interpretation. Check the methods and results, not only the abstract.
If the uploaded document is an abstract rather than a full paper, say so. A tool cannot extract a detailed method that is absent from the supplied material without bringing in another source, and that additional source should be identified rather than silently blended into the answer.
For critical appraisal, document retrieval is a starting point. Questions about bias, applicability and certainty require review of the study design and the evidence, not just a summary of the authors' conclusions.
Use page references as an audit route
Doximity's published description qualifies page and section references with "when applicable". Do not upgrade that into a promise that every statement always has a page citation.
Where a reference appears, open it and check the relevant passage. Confirm that the page numbering corresponds to the document you are using: a PDF viewer's page count and a printed page number may need to be distinguished during manual review.
Record unsupported answers, omissions and source-location problems separately. A correct statement with an unusable reference presents a different review problem from a confidently incorrect extraction.
No error rate can be estimated from the fictional examples in this article. A real evaluation would require actual outputs, a defined denominator and a reproducible review method.
Keep source content separate from new advice
A document may describe an old plan, a proposed plan or an action that actually occurred. A summary should preserve those distinctions before any further recommendation is considered.
Use explicit requests such as "What does the letter recommend?" and "What evidence would be needed to assess that recommendation?" as separate steps. Otherwise, a model may appear to answer the first question while actually supplying a new interpretation.
This article is published by iatroX and includes its clinical reference as a complementary tool. Per its September 2026 product specification, Ask-iatroX offers free source-linked answers grounded in NICE, CKS, SIGN and SmPC information from emc. That can help with a separate knowledge question after the document's facts have been checked.
It is not a claim that iatroX provides the PDF-upload workflow described here, reads the patient's record or resolves an unclear source document automatically. Keep document extraction and evidence checking as distinct jobs.
A defensible output is a reviewable output
For a long referral, use the tool to locate facts and discrepancies, then verify the source passages. For a laboratory report, preserve sample timing and result status. For a paper, distinguish the reported outcome from its interpretation.
The value of conversational document review is the possibility of reaching the relevant source more efficiently. The final standard remains whether the answer accurately represents the document you supplied.
Frequently asked questions
Does Doximity Ask support research PDFs as well as clinical letters?
Yes, its July 2026 announcement includes research papers and several clinical document types. The published description is not a guarantee that every file layout will be interpreted correctly.
Will every answer include a page reference?
The announcement says relevant page and section references are provided when applicable. Check the actual output rather than assuming universal citation coverage.
Should a missing result be interpreted as normal?
No. In a document-review task, an absent result should remain recorded as unavailable unless another identified source supplies it.