Two AI tools, one model provider, one rule: your document text is sent only when you click. Here's the data path, the prompts, and the limits.
Arawa PDF has two AI tools at launch — Summarise PDF and Translate PDF. Both are Premium. Both are powered by Anthropic's Claude. And both are opt-in: clicking the tool is the consent, and nothing leaves our infrastructure for the model provider unless you explicitly invoke them.
This post walks through what happens when you click "Summarise" or "Translate", and what doesn't happen.
When you upload a PDF and click Summarise, we run a fast text-extraction pass on the document. That extraction happens entirely on our servers using PyMuPDF; the PDF file itself never leaves our infrastructure.
We then send the extracted text (not the original PDF, not any embedded images, not any of your account metadata) to Anthropic's Claude API. The request includes the text and a short system prompt that tells the model what kind of summary or translation to produce. Nothing else.
Anthropic publishes that API inputs are not used to train its models. The text we send is processed only to generate the response, and is not retained for model training.
A few things that explicitly stay on our side of the line:
• The original PDF file. We send text, not files.
• Embedded images inside the PDF. The text-extraction pass doesn't extract them.
• Your account identifier, email, or name. The model has no idea who you are.
• Any data from other documents you've processed. Each call is independent.
It would be slightly easier to run Summarise automatically on every upload — pre-warm the cache, snappier UX. We deliberately don't. The model provider is one of the very small number of outside parties that ever touches your document content (the others are Stripe for billing and an SMTP provider for emails). We want "do they touch my data?" to have a yes-or-no answer that's tied directly to a button you clicked.
This means there's no Summarise running silently in the background. No invisible call. No "helpful suggestions" trying to guess what you might want next. You click, it runs.
There's a per-document size ceiling on AI tools — currently around 200 pages or 500,000 characters of extracted text, whichever comes first. Documents larger than that are summarised by chunking; the chunks are summarised independently and then collapsed into a single output. The model and prompt details are subject to change without notice, but the data-flow rules above are stable.
Rate limits exist on Premium: a small number of AI calls per minute, and a generous daily cap. If you hit them, the tool tells you so; nothing fails silently.
On the roadmap, in rough order: better handling of structured tables in Summarise output; per-paragraph translation that preserves layout more faithfully; a "chat with this PDF" tool that keeps the same opt-in / no-training data path. We won't add an AI tool that breaks the rules we've described here.
Questions, requests, or you found a case where the output is bad? support@arawapdf.com. We use Summarise and Translate ourselves and improve the prompts based on real-world failures.