What it does
Upload documents and ask questions about them in plain language. Supported formats include PDF, Word, PowerPoint, Markdown, plain text, HTML, JSON, XML, YAML, CSV, TSV and Excel, and scanned PDFs are read with built-in optical character recognition. The point is to get an answer that is anchored in your own material, not the open internet.
Every answer shows its sources
Answers carry numbered markers that open an evidence drawer showing the document, section, page and passage each claim rests on, a confidence score, and a label of whether the answer was generated or read deterministically. When no uploaded passage supports the question, the assistant says so and abstains instead of inventing an answer. Instructions hidden inside an uploaded document are stripped before the model sees them.
Chat with your database, safely
Connect a PostgreSQL, MySQL or SQL Server database and ask a question in plain language. The assistant drafts a read-only SELECT for you to review before it runs, and it executes in a read-only transaction with a statement timeout, a row limit and a host guard. You approve the query; the assistant does not run writes.
Export an answer
Any answer can be exported as Markdown, CSV, Excel, JSON or a chart. The export is rebuilt on the server from the stored, tenant-scoped answer, never from anything the browser sends back, and it carries spreadsheet formula-injection defences. An empty export returns an error rather than a blank file.
Private, and honest about being AI
- Your documents are private to your tenant, stored with row-level-security isolation, and every request takes your identity from your signed-in session, never from the request body.
- Answers are labelled as AI-generated and carry a confidence score. This is an assistant, not a system of record, and the interface keeps that framing as the EU AI Act transparency rules expect.
- Answers take several seconds, because the model runs on the platform on shared CPU. We would rather be honest about that than promise instant replies.
- Plan limits apply: Free 3 documents, Solo Pro 25, Team 100, Enterprise 500, with a fair-use daily question cap.
Questions people ask about this
- Do my documents go to an external AI provider?
- No. Answering runs on a local model on the platform. No document, prompt or answer is sent to OpenAI, Anthropic, Google or Meta, and no customer data trains any model.
- How fast is an answer?
- Several seconds. The model runs on shared CPU on the platform, so we do not promise instant replies. Accuracy and honest citations matter more here than speed.
Source: /documentation/document-intelligence, updated 2026-09-18. Answers are checked against the shipped product on that date.
Found something out of date? The people who build the module answer the contact form on the home page, and corrections to this guide are dated above. Tell us.