Extract key points from long documents with hierarchical text chunking
Reviewing 80-page contracts, financial statements, or academic papers often means choosing between slow manual skimming and uploading confidential files to cloud chatbots. This PDF summarizer avoids both trade-offs by running two distinct summarization engines directly inside your browser with 0 bytes transmitted over the network.
Chunking long PDFs for Chrome’s built-in Gemini Nano model
When AI summary is active, the tool uses the Chrome for Developers Summarizer API documentation implementation built into desktop Chromium browsers, powered by Chrome’s built-in Gemini Nano model. Because on-device language models have finite context windows, any document exceeding 9,000 characters is split along paragraph boundaries automatically.
Each chunk is summarized separately, then combined in up to 4 recursive merge rounds until the distilled notes fit a single final pass. If a model call still hits an input limit on dense text, it halves the block and retries cleanly down to an 800-character threshold without failing the whole job.
Ranking verbatim sentences to prevent hallucinations
Key sentences works in every desktop and mobile browser with zero model downloads and helps prevent hallucinations when reading financial statements or legal contracts. It strips common stopwords, builds TF-IDF vectors for every sentence containing at least 4 content words and up to 400 characters, and scores each sentence by cosine similarity against the document.
Opening sentences receive a 15% position boost to capture executive introductions, while near-duplicate sentences above 0.8 cosine similarity are skipped. If you need the full unedited body copy before or after summarizing, you can extract raw unformatted text first.
When should you pick AI summary vs. key sentences?
Pick AI summary when you want synthesized bullet points or a one-line headline on desktop Chrome or Edge, and pick Key sentences when you need verbatim quotes with page numbers in any browser. Matching the mode to your accuracy and browser requirements avoids wasted time.
| Feature | AI summary | Key sentences |
|---|---|---|
| Browser support | Desktop Chrome or Edge | Every desktop and mobile browser |
| Output style | Rewritten prose or bullets | Exact source sentences with page numbers |
| Hallucination risk | Low, but requires fact-checking | Zero, quotes verbatim text |
| Model download | One-time browser download | None, runs immediately |
| Configuration options | 3 styles and 3 lengths | 5, 8, 12, 20, or 30 sentences |
In AI summary mode, the tool works best on English, Spanish, French, German, and Japanese documents. Key sentences supports any language that separates words with spaces and punctuation, making it reliable for multi-lingual technical reports.
Behind the scenes, exporting your results to Word runs off the main UI thread using the MDN Web Workers API documentation pattern so the browser tab stays responsive. For preserving full headings and tables instead of a short summary, try structured Markdown document export.
Handling scanned PDFs, pasted text, and export formats
Documents arrive in many forms beyond clean digital PDFs. Whether you drop a scanned paper archive or paste meeting notes directly into the editor, the summarizer normalizes the input into clean paragraphs before scoring or condensing.
Running automatic OCR on image-only pages
Under the Library of Congress PDF 1.7 (ISO 32000-1) format description, scanned PDFs store raster page images without character streams. When the reader detects pages with little or no embedded text, it runs local OCR automatically and lists which page numbers were scanned.
At least 50 characters must be extracted from a PDF, or 200 characters entered in Paste text, before summarization starts. For foreign-language PDFs that you need to read in your own language first, you can translate foreign-language reports privately on your device.
Exporting summaries to Word (.docx), plain text, or clipboard
Once your summary appears, Copy writes clean bulleted text straight to your system clipboard. Selecting Word (.docx) packages the heading and bullet points into a formatted .docx file inside a Web Worker, while Text (.txt) saves lightweight UTF-8 plain text with zero watermarks or formatting bloat.
If a document contains 50 or more pages of dense tables, reviewing the extracted page citations beside each key sentence helps you jump straight to the exact source page in your viewer. You can also switch between Style and Length options and select Summarise again without re-uploading or re-parsing the PDF.
Why is AI summary unavailable in some browsers?
AI summary is disabled or unavailable when a browser lacks Chromium’s built-in on-device language model, runs on a mobile phone, or lacks free disk space for the model cache. Key sentences remains active in every browser so you can still summarize immediately.
Desktop hardware, storage, and first-run model downloads
Desktop Chrome and Edge require Windows, macOS, Linux, or Chromebook Plus hardware with at least 22 GB of free storage and sufficient system RAM before caching the local model. On your first run, the browser displays a percentage progress indicator while downloading the model locally for future offline use.
Tip: If Chrome reports that a browser gesture is needed to start downloading the model, select Summarise a second time directly with your mouse or keyboard so the browser permits the download.
Using Key sentences as an instant cross-browser fallback
If you work in Firefox, Safari, or on a phone, Key sentences selects 8 representative sentences by default and lets you expand up to 30 sentences for long legal or technical briefs. Afterward, you can listen to your summary read aloud using your device’s local voices.