Sent raw
the whole document, straight to the model
Playground
One question, one model, asked twice — once with the whole document, once with only what your question needed. Same answer, a fraction of the tokens. No account, nothing to install.
Plain text only — .txt, .md, .csv, .json, .log. No PDF, Word or Excel, on purpose — here's why.
The question is the ruler. Everything in the document is scored against it, and a specific question protects more than “summarise this”.
the whole document, straight to the model
only the spans your question needed
Timings are from the run you just made. The last line is the point the request buffer is released — after it, the only trace of this run anywhere is a row of numbers used for the rate limit. Privacy Policy →
Language models make mistakes on their own, with or without compression, and both columns here come from the same one. Read the two answers against each other — that comparison is what this page measures. Neither column is a fact-checked source.
The demo runs a small long-context model on our key so you don't have to bring one. With your own token you can run the same test against any model your provider offers — GPT, Claude, Gemini, or something you host yourself.
A narrow question keeps a little; “summarise everything” keeps a lot. Both are correct behaviour. Try the same document with two different questions and watch the numbers move. See the benchmarks →
Because the safest document parser is the one that doesn't exist. PDF, Word and Excel files are containers: readers for them unpack embedded objects, fonts, scripts, macros and links to outside resources, and that machinery is among the most reliably exploited software there is. We run none of it. A file you pick here is read in your browser and only its text is sent — the file itself never reaches our servers, so there is nothing on our side for a malformed document to act on.
It protects you as well as us: with plain text you can see exactly what you are sending. Nothing is hidden in a container you can't read. Anything that doesn't look like text is refused, in the browser and again on the server.
Create an account, generate a token, and point your existing SDK at CompresLM. Your provider key stays yours, your model stays your choice, and there is no limit on how much you test.