What is a Text Summariser?
Summarising a document takes one of two shapes. An extractive summary chooses sentences that already exist in the source and stands them next to each other, in their original order and their original wording. An abstractive summary — what an AI model writes — produces new prose that merges, reorders and paraphrases what the document said. Both are legitimate; they fail in opposite directions, and knowing which you are using is the whole point.
The extractive method here is a well-understood scoring exercise, not machine learning. Words are counted across the document, rare words are discounted, and each sentence is scored by the words it contains relative to how common they are overall — with a small penalty for length so that a single rambling sentence does not dominate. The highest-scoring sentences are then printed in their original order. Nothing is invented, and every line can be traced back to the source, which makes it the safer method whenever the summary will be checked against the original.
The AI mode calls a chat-completions endpoint with a key you supply. Nothing is proxied through this site: your browser talks to the provider directly. That produces readable prose the extraction method cannot match, and it introduces a real risk — models state things the source never said, especially where the document is thin, and a rewritten summary gives you no way to tell which sentence came from where. The tool is built so you can compare the two outputs on the same text and decide for yourself.
How to summarise a long text
- Paste the article, report, transcript or meeting notes into the source box. The word and sentence counts below it tell you what you are working with.
- Start in on-device mode and set the ratio you want to keep. Thirty per cent of the sentences is a reasonable default for a report; ten is closer to an abstract.
- Read the keyword chips under the result. They are the words the scoring rewarded — if that list looks wrong, the summary will read wrong, and you will see why.
- Adjust the ratio until the summary carries the argument without the repetition.
- If you want readable prose rather than stitched sentences, switch to AI rewrite, paste your own provider key and set what you want back.
- Compare the two on the same document before deciding which to trust for the job at hand.
Text Summariser limits and rules
| Item | Limit |
|---|---|
| On-device method | Extractive scoring — sentences are selected, never rewritten |
| Sentence ratio | 10% to 70%, in steps of five, defaulting to 30% |
| Shortest useful input | Two sentences; below that there is nothing to select from |
| AI mode | Any provider speaking the OpenAI chat-completions shape, including a local server |
| Key storage | Only if you tick the box — otherwise it lives in memory for the tab and is gone when you close it |
| Cost | On-device mode is free and offline; AI mode is billed by your provider, not by this site |
| Handling of your text | On-device mode sends nothing; AI mode sends the whole document to the provider you nominate |
| Languages | Any; the sentence splitter also breaks on the Devanagari and Bengali danda |
When to use it
- Cutting a research paper down to enough to decide whether to read it fully.
- Turning an hour-long transcript into notes you can actually find something in again.
- Reducing a long email thread to the decision it reached.
- Getting the gist of a contract clause before sending it to somebody qualified.
- Making a first pass at an abstract that you will rewrite by hand, with the made-up parts removed.
- Summarising confidential material when uploading it to a service is not an option.
Is it private?
In on-device mode, nothing leaves the page: the scoring runs as arithmetic on your text and there is no network request at all, which you can confirm in the network inspector. In AI mode your browser sends the document directly to the provider whose key you supplied — this site is not in that path and sees neither the text nor the key. If you tick the remember box, the key is written to this browser's localStorage, where any script on this domain could read it; leaving it unticked keeps the key in memory for the tab only. The text itself is never persisted by either mode.
Frequently asked questions
An extractive summary is built from sentences that already exist in the source and keeps their wording, so nothing can be invented and every line is traceable. An abstractive summary, which is what an AI model writes, is new prose that can merge and paraphrase — it reads far better but can state things the source never said and gives you no way to check which parts came from where. This tool offers both so you can compare them on your own document.
No. The default mode runs entirely on your device using sentence scoring, needs no key, no account and no network connection once the page has loaded, and costs nothing. The key is only required if you switch to AI rewrite, which is an optional way to get smoother prose from a provider you already pay for.
Your browser sends it directly to the provider's endpoint as part of the request; it is not sent to this site, because there is no server side to this tool. If you tick the remember box, it is written to this browser's localStorage, where any script running on this domain and anybody with access to this computer could read it. Use a key with a spending limit and a short expiry, leave the box unticked on a shared machine, and remember that on-device mode needs no key at all.
Because it is selecting sentences, not writing them. The connections between ideas live in the sentences that were dropped, so the result can feel like a list of statements rather than an argument, and unusual word choices from the source survive intact. That lumpiness is the price of a summary that cannot invent anything. If you need it to flow, use AI rewrite — and check the result against the source.
Every word is counted across the document, and words that appear everywhere or almost nowhere are discounted. Each sentence is then scored by the frequency of the words it contains, normalised by its length so that one long rambling sentence does not win, with a small bonus for the opening sentences because they usually frame what follows. The highest-scoring sentences are printed in their original order. It is ordinary arithmetic and it is entirely predictable once you know the rule.
Yes. The AI mode accepts any base URL that speaks the OpenAI chat-completions shape, so a server running on your own machine — the usual local model runners all expose one — works the same way. Point the base URL at localhost, put anything in the key field if your server does not require one, and the document stays on your computer.
Related tools
Written and maintained by Md. Nur Hossain Repon.