What the interface processes.

Conversation text travels through Vercel and an authenticated gateway to Qwen running on Modal's external GPU infrastructure, operated by psr.ai. When the local fallback is configured and Modal is not ready, requests instead travel through ngrok and an authenticated gateway to a model on the owner's computer. The model and precision can differ between these backends. Requests already in progress are not replayed on the other backend.

TXT, Markdown, PDF, Word, and Excel attachments are parsed in a short-lived worker in your browser. Original file bytes do not leave the browser. The filename, format, truncation status, and bounded extracted text then follow the same request path as the conversation. Legacy DOC extraction is beta.

The application does not write prompt, response, filename, or extracted document text to its database or analytics events, and submitted content is not used by psr.ai to train a model. Vercel, Modal, and ngrok (when using the local fallback) process content in transit and memory while serving requests and may retain network and operational metadata under their own policies. Readiness polling contains no conversation or document text. The application does not queue prompts in Modal functions or store them in model caches.

Abuse protection derives a keyed, one-way pseudonym from the request IP. The raw IP is not stored in the rate-limit database, and those counters expire within 24 hours.