You open an 80-page report, a catch-all contract or a technical brief, and you need the essentials without copy-pasting it into ChatGPT. AI embedded directly in the browser promises exactly that: a summary generated on your machine, without the document ever leaving your device. It works — but under precise conditions that this article details.
How local summarization works
Document Flare’s summarization tool relies on the Summarizer API built into Chrome and Edge from version 138. The browser downloads a small language model once — typically Gemini Nano or Phi depending on the browser configuration — then runs inference locally. Your PDF stays in the browser: its text is extracted page by page with pdf.js, split into chunks of about 15,000 characters, then summarized by a map-reduce mechanism that condenses each chunk before producing the final result.
No HTTP request carries your document. The only network traffic concerns the initial page load, its resources and the AI model on first use. This is exactly what distinguishes local processing from cloud processing. To understand how to verify it yourself, see our 5-minute network test (article available in French).
Browser requirements
Two conditions are required:
- Chrome or Edge version 138 or later. The Summarizer API is not implemented in Firefox, Safari, or in most Chromium-based browsers.
- The AI model must be available for your browser. Even with the right version, the API may be inactive depending on your region, your corporate policy or experimental flags.
Check your version
- Open Chrome or Edge.
- Click the ⋮ menu at the top right, then Help > About Google Chrome (or Microsoft Edge).
- The browser displays the version number and updates itself automatically if it’s not current.
If the displayed version starts with 137 or lower, the tool will show an unavailability message and won’t be able to run the summarization.
Step by step: get a summary
- Open the Summarize a PDF with AI tool.
- Drop your PDF. The tool accepts a single file, up to 50 MiB and 1,000 pages.
- Choose the summary type. Two options are offered: key-points for a structured list, or tl;dr for a condensed paragraph.
- Choose the length. Short, medium or long depending on the level of detail you want.
- Start the processing. The first use may trigger the model download; a progress bar shows its progress.
- Retrieve the result. Once the summary is generated, you can copy it or download it as Markdown.
What you get: Markdown text, not a transformed PDF
The result is a Markdown text containing the summary. The source PDF is neither modified, nor annotated, nor replaced. You can copy the text into a word processor, archive it in a note or reuse it in another document. It’s a raw textual output: the PDF layout, tables and images are not preserved.
This limitation is deliberate: reinjecting a summary into a PDF would mean rewriting the document, which is not the tool’s purpose. If you’re looking for a structured text version of a PDF, the PDF to Markdown tool remains more suitable.
Honest limits to know
Very restricted browser availability
The Summarizer API is an experimental Chrome and Edge feature. If you use another browser, the tool will simply display an unavailability warning. There is no workaround: the model is provided by the browser itself.
First model download
The first use downloads the AI model locally. Depending on your connection, this can take a few seconds or a few tens of seconds. Once cached, subsequent uses are noticeably faster. Offline and without a prior cache, the tool is unavailable.
Quality varies with the document
The summary depends on the quality of the extracted text. A PDF with a complex layout, poorly ordered columns or non-standard fonts can produce approximate text before the AI even steps in. The embedded model is also less powerful than large remote models: summaries may be terse or lack nuance on technical subjects.
Requires a text layer
The tool extracts the PDF’s native text. A scanned document or an image-only PDF with no selectable text will return an error. You first need to run a local OCR to obtain the text.
Size and pagination limits
The capability catalog sets the limits at: a single file, 50 MiB maximum, 1,000 pages maximum. PDFs beyond these limits are rejected before processing.
Language and consistency
The summary is produced in the detected language of the source text, but quality varies with the language and the installed model. Don’t expect automatic translation: the tool summarizes, it doesn’t translate. To translate, use the Translate a PDF tool instead, which relies on the same local infrastructure.
Verify that nothing goes to a server
The promise of local processing can be verified concretely. Open the developer tools (F12), go to the Network tab, then drop your PDF and run the summarization. You should observe no POST or PUT request carrying the document. The only outgoing requests concern the page load and, on first use, the AI model download. For a complete method, follow our network test protocol (article available in French).
To go further on what local processing actually protects — and what it doesn’t — read PDF privacy: what local processing really protects.
When to use this local AI summary
This tool is relevant when:
- you want a quick synthesis without sending a confidential document;
- you already have Chrome or Edge 138+;
- the PDF contains native text and fits within the size limits;
- you accept a Markdown text output rather than a transformed PDF.
It doesn’t replace a remote model for very fine-grained analysis, nor OCR for scans. Its main advantage is keeping the document on the device.
If you prefer to query the document rather than get a summary, the Semantic search tool explores the text with locally computed embeddings, on the same principle of a model downloaded once.
