Skip to main content
Free plan, no cardStart free

DOCDocument intelligence · updated 2026-11-14

AI document assistant: answers cited to your own files

Upload PDFs, Word, Excel, Markdown and more and ask questions; answers cite the document, section and page, abstain when unsupported, and come from a local model by default.

For: QA leads and product owners who keep specifications, test plans and policies as documents and want answers they can check, security and compliance teams that need the citation, and anyone who wants to ask a database a question without writing the query.

acme-shop / checkout-flow.recIllustrative
1 selector healed#pay-now-btn[data-test="pay"]

Capabilities

Your formats, including scanned PDFs

PDF, Word, PowerPoint, Markdown, plain text, HTML, JSON, XML, YAML, CSV, TSV and Excel; a scanned PDF is read by optical character recognition before it is indexed.

Read more: Your formats, including scanned PDFs

Every answer shows its sources

Numbered markers open an evidence drawer with the document, section, page and passage each claim rests on, a confidence score, and a label saying whether the answer was generated or read deterministically.

Abstains rather than invents

When no uploaded passage supports the question, the assistant says so instead of answering. Instructions hidden inside an uploaded document are stripped before the model sees them.

Ask your database, read-only

Connect PostgreSQL, MySQL or SQL Server and ask in plain language. The assistant drafts a SELECT for you to review, then runs it in a read-only transaction with a statement timeout, a row limit and a host guard. It does not run writes.

Export an answer

Markdown, CSV, Excel, JSON or a chart, rebuilt on the server from the stored answer rather than from anything the browser sends back, with spreadsheet formula-injection defences.

Private to your tenant

Documents are isolated per tenant with row-level security, every request takes your identity from your signed-in session, and answers are labelled as AI-generated with a confidence score.

How it works

  1. UploadThe documents you want answered, within your plan’s allowance; a scanned PDF is read by OCR.
  2. AskA question in plain language. The answer arrives in several seconds, with its markers.
  3. Check the citationOpen a marker to read the passage, the page and the confidence; an abstention tells you nothing supported the question.
  4. Export, or ask a databaseExport the answer as a file or a chart, or connect a database and approve the SELECT the assistant drafts.

What you receive, and what we do not do

Evidence ledger
Per answer
The passages cited (document, section, page), a confidence score and the generated-or-deterministic label.
Per database question
The SELECT drafted, your approval, and the timeout and row limit it ran under.
Per export
A file rebuilt from the stored, tenant-scoped answer; an empty export is an error, not a blank file.

Limits we state

  • On the local model, answers take several seconds: it runs on shared CPU inside the platform, and we would rather say so than promise instant replies.
  • Plan limits: Free 3 documents (10 MiB each, 3 per question), Solo Pro 25 (25 MiB, 5 per question), Team 100 (50 MiB, 10), Enterprise 500 (100 MiB, 10), with a fair-use daily question cap per tenant.
  • The database assistant drafts and runs SELECT statements only, each shown to you first; it never runs a write.
  • It is an assistant, not a system of record: a person reads the citation and decides.

Plans: On Free as the Document intelligence single tool, $19 a month, with 3 documents; on Solo Pro as an add-on with 25; included on Team with 100 and on Enterprise with 500. See the metering table.

Questions buyers ask

Do my documents go to an external AI provider?

No. Answering runs on a local model inside the platform; this feature is not one the Work with AI add-on covers, so no document, prompt or answer is sent to an external provider, and no customer data trains any model.

How is this different from asking a chatbot?

The answer is anchored in your files: every claim carries a marker to the passage it rests on, the confidence is shown, and when nothing in your files supports the question the assistant abstains instead of guessing. A chatbot over the open web does none of those three.

Can it change data in my database?

No. It drafts a SELECT, shows it to you, and runs it only after you approve, inside a read-only transaction with a statement timeout and a row limit. Writes are not possible through it.

What does an answer look like when the files do not contain it?

An abstention: the assistant says that nothing in the uploaded documents supports the question. That is deliberate; an invented answer with a citation would be worse than none.

Try it on your own application

The Free plan asks for no card and runs for 14 days after the trial. Or bring one flow from your product and we record it live.

  • Free plan, no card
  • 14-day Solo Pro trial
  • Demo replies within one business day
  • Data hosted in India