✨ AI PDF Summarizer
Paste extracted PDF text or notes, or drop a PDF. PDFSwift will summarize it with the site AI.
A short completion of the text the tab could read
The summarizer reads a PDF in the browser first. Text it finds is sent with task: "summary" to /api/ai. The Worker keeps the first 30,000 characters, builds a prompt that asks for an executive summary, key points, and important details, and runs the summary model list with max_tokens: 1800 and temperature: 0.3. The first model that returns text wins. The response is shown in the page. You copy it. It is not written back into the PDF.
HTTPS covers the POST. The Worker code does not persist the prompt. The model provider still receives the text for that call. A client file you are not allowed to send does not become allowed because the button says summarize. Split to the chapter you need. Paste a note in the box if the PDF text is messy, and label which part is the note so it is not summarized as if the author wrote it.
| Summary task | |
|---|---|
| Endpoint | POST /api/ai |
| Text cap | 30,000 characters |
| Answer cap | about 1,800 tokens |
| Empty scans | Nothing useful until OCR |
Briefs that still get checked
A 40-page report, one meeting
Summarize the findings chapter, not the appendices. Open the PDF and confirm each number in the bullets. A rounded figure that the model “cleaned” is a different figure.
A thread of emails exported as one PDF
The text layer may be in a strange order. If the summary attributes a sentence to the wrong person, the extract was interleaved. Fix the extract or summarize a shorter slice.
A scan of minutes
OCR, fix names, then summarize the corrected text. Summarizing the picture directly is not what this task does.
A decision you have not made
If the output says “the team should,” and the paper only listed options, delete the sentence. The prompt asked for details. It did not appoint you.
Read the source against the bullets
- Drop a text PDF or paste text. If highlight-select fails in a reader, OCR first.
- Generate. Copy the result into a note, not into the official file.
- Check every number and every name against the PDF.
- If the document was longer than the cap, summarize the next slice and do not merge the two notes until you have labeled which pages each one covers.
Summarizing a long PDF is the case where the recommendation was not in the paper.
Two slices, two labels
When the report exceeds 30,000 characters, the second half is not summarized at all. Run a second summary on the next chapter and title the note with the page range. A combined note that hides which half it came from will mix a finding from chapter 2 with a number from the appendix. The completion cap, about 1,800 tokens, also means a “full” summary of a long paste is a short summary of the start.
Frequently asked questions
Does the summary include pages that are only pictures?
Only if those pages yielded text. A scan is extracted as images and, on this task, the text layer is what gets summarized. An empty layer produces a summary of nothing, or of the caption you typed. OCR the pages first and paste the checked transcript.
Why did the back half of a long report disappear?
The Worker slices text at 30,000 characters before the model sees it. The completion is capped around 1,800 tokens at temperature 0.3. Split the report and summarize each part. A single button cannot hold a 200-page brief.
The summary added a recommendation that was not in the paper.
The system prompt asks for an executive summary, key points, and important details. It does not ask the model to invent a decision. Still, models pad. If a sentence is not in the source, delete it. Do not send the summary as the paper.
Is the document stored?
The text is posted as JSON to /api/ai. The Worker does not write that body to a database. It forwards it to the first summary model that returns text and sends the completion back. Do not summarize a file you are not allowed to transmit.
Can I ask a follow-up here?
This page is the summary task. A question about one clause belongs on AI PDF Chat, which sends the question separately and is instructed to say when the document does not contain the answer.
The output looks confident and the source was a scan I did not OCR.
Confidence is not evidence. Open the PDF at the page the bullet claims to describe. If you cannot find the sentence, the bullet is not a summary.