Abstractive vs extractive: why this matters for your results
There are two fundamentally different ways to shorten text, and knowing which you’re using sets the right expectations:
| Extractive | Abstractive (this tool) | |
|---|---|---|
| Method | Picks the most important existing sentences | Writes new, shorter sentences |
| Reads like | A highlighter pass | A human paraphrase |
| Faithfulness | High, every word is from the source | High but not perfect, can rephrase imperfectly |
| Risk | Choppy, disconnected | Rare hallucinated detail |
This tool uses abstractive summarization (a DistilBART model), which is why the output reads smoothly rather than like disconnected highlights. The trade-off is the small hallucination risk above, a fair price for summaries that actually sound written rather than stitched together.
Everything runs on your machine, which has consequences
Because the ~230MB model executes locally in a Web Worker, two things follow. First, privacy is absolute: your text never leaves the browser, so it’s safe for confidential drafts, contracts, and proprietary research. Second, speed depends on your hardware: a fast laptop summarizes in a few seconds; an older device or a phone takes longer, and the first run also has to download the model. This is the deliberate trade: no servers, no uploads, no per-request cost, in exchange for using your own compute.
Getting cleaner summaries
- Feed it clean prose. Strip navigation text, ads, and reference lists before summarizing, the model treats everything as signal and will dutifully summarize the noise.
- Chunk long documents. Summarize each major section, then combine. This avoids the mid-document context loss that hurts single-pass runs on long inputs.
- Mind the output limit. Transformer models cap their output length, so a summary of a very long text can end abruptly. If it cuts off mid-thought, your input was too long for one pass, split it.
When to reach for something else
A summarizer compresses; it doesn’t analyze, fact-check, or answer questions. If you need a specific answer buried in a document, the AI Question Answering tool extracts it directly. If you need to pull out names and places, the AI Entity Extractor is the right instrument. Use the summarizer for the “what is this about, quickly?” job it does best.
Summarizing a PDF
The Summarize a PDF mode above, or jump straight to it.
A two-stage pipeline, entirely in your browser
PDF summarization is really two jobs chained together, and both run locally on your device. First, PDF.js extracts the text from your document. Then the DistilBART model (via Transformers.js) reads that text and writes an abstractive summary. Nothing uploads, your contracts, research papers, and reports are parsed and summarized on your machine. That makes this safe for exactly the confidential documents you’d never paste into a cloud summarizer.
Text PDFs vs scanned PDFs
This distinction trips up more people than any other, so it’s worth internalizing:
| Text-based PDF | Scanned PDF | |
|---|---|---|
| Made by | Exporting from a document app | Scanning/photographing paper |
| Can you select text? | Yes | No |
| This tool can read it? | ✅ Directly | ❌ Needs OCR first |
If summarization returns nothing or gibberish, check whether you can highlight text in the original, that one test tells you whether you have a text PDF or a scan that needs the OCR detour.
Handling long documents
Transformer models have a fixed context window, so a 100-page report can’t be summarized in a single pass. The tool processes long text in chunks and combines the results, which works well but can lose some connective thread between distant sections. For very long or highly structured documents, you’ll get a more coherent result by summarizing the key sections separately. And because processing is local, a long document’s speed depends on your hardware; expect a wait on large files.
What a good summary gives you: and what it doesn’t
Use the summary to answer “what is this document about and what are its main points?”, its real strength. Don’t rely on it for exact figures, the contents of tables, or fine legal/contractual specifics; for those, the summary points you to the right section and you read the original. To pull specific facts out of the document instead of an overview, the Question Answering tool extracts precise answers grounded in the text.