Explixit
PricingDownload Chrome Extension
Back to product

AI & THIRD-PARTY PROCESSING

AI and Third-Party Processing

What enters the AI pipeline, how much context each feature uses, and how to interpret model output.

Version date: 1.1Updated: 2026-07-22
Configuration: Requests that require paper context use a third-party Responses-compatible provider and include the complete PDF in each request. Minimized text features may use a separate text-model gateway. This page will be updated when the model or supplier changes.

1. AI processing flow

  1. 01The extension keeps the current PDF in the browser and calculates its content hash.
  2. 02Each document AI request includes the complete PDF. The API verifies the actual-byte hash and ownership.
  3. 03The API loads conversation history from PostgreSQL and organizes the selection, page number, or local PDF.js segments.
  4. 04The complete PDF, history, and current request are sent to the configured stateless model provider.
  5. 05The response returns to the reader. The API stores permitted results and usage, then releases the request's PDF bytes.

2. Content used by each feature

  • Summary and keywords: use the complete original PDF to generate and cache a document-level overview.
  • Paper chat: sends the complete PDF, server-side conversation history, and current question on each turn.
  • Selection explanation: sends the complete PDF, selected text or formula, physical page number, and normalized coordinates.
  • Selection translation: translates only the content you selected and does not expand it into a full-paper request.
  • Page translation: sends the complete PDF and locally generated PDF.js segments for the current page, and stores per-page state and results.
  • Automatic highlights: uses the complete PDF to identify methods, contributions, results, or other selected categories.
  • Visual explanation: when enabled, sends the confirmed page crop pixels, page number, region, and complete PDF.

3. Models and suppliers

The client selects only server-published aliases and cannot submit an arbitrary provider model ID. The server maps aliases to models verified against the actual protocol. Provider support for Files APIs, file reuse, sessions, or other unverified behavior must not be assumed.

The document provider is treated as stateless, so the same PDF may be sent again in later requests. Suppliers process requests under their own terms. Without contractual and configuration evidence, Explixit does not promise zero retention or that data is never used for training.

4. Formulas and rich text

Explixit tries to preserve formula structure and renders model answers with Markdown, DOMPurify, and local KaTeX. A model may still omit variables, subscripts, transposes, or equation conditions. Verify important formulas against the PDF.

5. Visual explanation

A visual-explanation request should contain a PNG or JPEG cropped from actual PDF canvas pixels, rather than only OCR, a caption, or fixed test data. The server validates ownership, page number, MIME type, file signature, size, and pixel dimensions.

Explixit does not persist the cropped source image. It stores only necessary source metadata and the model answer. Processing remains subject to provider and infrastructure logging policies.

6. Output limitations

  • Models may produce factual errors, incorrect citations, wrong formulas, or conclusions that are not in the paper.
  • Scan quality, encryption, file size, page count, or context limits may make a PDF unavailable.
  • The same question may receive different answers with a different model, time, or context.
  • Highlights are reading suggestions, not peer review or a judgment of paper quality.
  • Verify important material using the original PDF, reliable databases, and professional judgment.

7. User control

  • You decide whether to enable Explixit for a PDF.
  • Automatic page translation can be paused and resumed; one failed page does not clear other pages.
  • Streaming AI work can be cancelled; failure or cancellation must not appear as a completed result.
  • Manual and automatic highlights, comments, notes, and explanation records can be reviewed and deleted.
  • You can leave reading mode and restore Chrome's native PDF reader.

8. Security and feedback

API keys remain server-side. Error interfaces should not reveal keys, internal stacks, or fabricated fallback answers. Report model, infringement, or security issues to explixitai@outlook.com.