Documents

Read, compare and transform documents inside your workflows.

Seven workflow steps work with Word, Excel, PowerPoint, PDF, Markdown and HTML files right in the run: read them into text and rows, redline two versions, convert and split, pull values into a data table, remove personal data. No document service to buy or host — scanned PDFs and pixel-perfect Office to PDF are optional add-ons.

Document actions guide Four how-to guides

The document actions

Seven steps, one palette group.

Find them under Operations in the designer, next to the Word, Excel and PowerPoint template steps that fill documents in.

doc_read

Read Document

Any supported file into text, Markdown, sanitised HTML, metadata, structure — headings, tables, sheets, slides, comments, tracked changes, form fields — and rows to loop over.

Docs doc_compare

Compare Documents

Two versions of a document, spreadsheet or JSON: changes, statistics, similarity, a side-by-side HTML report and a Word or PDF redline.

Docs doc_convert

Convert Document

To PDF, Word, HTML, Markdown, plain text, CSV, Excel or JSON. Optional external converter for pixel-perfect Office to PDF.

Docs doc_pdf

PDF Tools

Merge, split, extract pages, count pages, watermark or stamp text, rotate, and fill or read form fields.

Docs doc_extract

Extract from Document

Tables, named values by pattern, and emails, URLs, phone numbers, dates, amounts and percentages — straight into a data table if you like.

Docs doc_redact

Redact Document

A copy without emails, phone and card numbers, IBANs and other personal data, or the words and patterns you choose.

Docs doc_inspect

Inspect Document

The real file type from the content, checksums, word and character counts, reading time, and search with context.

Docs ocr

Scanned PDFs and OCR

Pages without a text layer are found and recognised with tesseract.js, offline, or Azure AI Document Intelligence.

Docs

Formats

The files your business actually runs on.

The type is detected from the content — magic bytes, the parts inside a ZIP, text sniffing — not from the name. A mismatch is reported as a warning.

Text, Markdown and HTML in {{document.text}}, rows in {{document_rows}}
Metadata: title, author, dates, pages, word counts
Limits you set per instance: size, unpacked size, pages, time

Supported formats Limits

What Read Document reads from each format
Format	What you get
Word .docx	Text, headings, lists, tables, links, comments, tracked changes
Excel .xlsx	Every sheet, typed values, formula results, header detection, ranges
PowerPoint .pptx	Slide titles, text and speaker notes
PDF	Text per page, outline, links, form fields, tables; scanned pages with OCR
OpenDocument	.odt, .ods (as sheets), .odp
Markdown	Including front matter
HTML	Scripts and active content dropped
CSV, TSV	Delimiter detected, quoted fields
JSON, XML, YAML	Parsed into structure; a list of objects becomes rows
RTF and text	Text
Images	PNG, JPEG, GIF, WebP, BMP, TIFF with OCR

Compare

Redline two versions. Send the changes.

Compare contracts word by word, spreadsheets cell by cell with rows matched on key columns, or JSON with the path of every change. Get a Word redline with real tracked changes, a PDF redline and a side-by-side report ready to email.

Compare Documents settings: original and revised contract, compare by words, Word redline with tracked changes — Two contract versions, by words, with a Word redline in tracked changes.

Side-by-side comparison report in an email: deleted words struck through in red, inserted words underlined in green — The side-by-side report, used as the body of an email.

Compare Documents How to: contract redline

In and out

Attachments in, documents out.

Every document field takes a file object, a list of files, a storage reference, a URL or inline content. Every output is a file object the next step can use.

1
In
- A form's Attachments field
- A User Task upload
- HTTP Request and Download File
- Storage connectors: S3, Azure Blob, SharePoint, OneDrive, Box, Google Drive
2
In the run
- Read, compare, convert, split, stamp
- Extract values and tables
- Redact personal data
- Loop over rows, branch on what was found
3
Out
- Email with Send Email attachments
- Upload to storage or post to an API
- Insert rows into a data table
- Give the text to an AI agent

The Data panel listing Read Document's outputs: text, markdown, html, rows, metadata and structure

Every output in the data picker.

Read Document’s text, Markdown, rows, metadata and structure appear with their fields, ready to drag into the next step. Run history shows a compact summary, never the whole document.

Collect files with a form, read them in the next step
Summarise or classify with a Run Agent step
Write extracted rows to a data table, up to 5,000 a run

Read, then ask an agent.

There is no separate AI document step: Read Document turns the file into Markdown, and a Run Agent step with declared outputs summarises it, classifies it, and lets the workflow branch on the answer.

Summarise or classify with AI AI agents

json

[
  { "type": "doc_read", "label": "Read contract",
    "config": { "source": "{{form.contract}}", "max_chars": 60000,
                "output_var": "document" } },
  { "type": "run_agent", "label": "Summarise",
    "config": { "agent_name": "Contract reviewer",
      "user_message": "Five bullet points, then low, medium or high risk.\n\n{{document.markdown}}",
      "output_variable": "review" } }
]

Security

Built for files you didn't write.

Documents come from customers, suppliers and the internet. The readers assume the worst.

ZIP bombs

Office files are ZIP containers: the reader checks the directory first and inflates every part with a hard output cap.

No XXE

XML with a DOCTYPE is refused; only the predefined entities expand, with bounded depth and node count.

Sanitised HTML

HTML output is rebuilt from escaped text and passed through an allow-list sanitiser: no scripts, handlers or javascript: links.

Safe PDF parsing

pdf.js with eval off and no system fonts, so a crafted font cannot run code. Encrypted PDFs are refused clearly.

Your workspace only

Files are read and written for the run's workspace; another workspace's file is simply not found. URLs go through the egress guard.

Nothing leaves

The actions run in the app's own process with JavaScript libraries. Files only leave if you configure cloud OCR or an external converter.

Security details How uploaded files are kept safe

How-to guides

Four workflows to copy.

Test run logs: Read Document read three rows from two sheets of a workbook, then a loop logged each order

Save & Test: Read Document reports the format, sheets and rows; the loop runs once per row.

Questions

Which file formats can a workflow read?

Word, Excel and PowerPoint (.docx, .xlsx, .pptx), OpenDocument, PDF (scanned pages with OCR), Markdown, HTML, CSV and TSV, JSON, XML, YAML, RTF, plain text, and images with OCR. The type is detected from the file's content, not its name. Older .doc, .xls and .ppt files convert to PDF with the optional external converter.

Do documents leave my infrastructure?

No. The document actions run inside the app with JavaScript libraries and need no external service. A file only leaves if you choose a cloud OCR provider or point the optional Office-to-PDF converter at a service — and that should be one on your own network.

Can it produce a Word redline with tracked changes?

Yes. Compare Documents can write a Word redline as real tracked changes (or as formatting: green underline for inserted text, red strike-through for deleted), a PDF redline, and a side-by-side HTML report you can use as an email body.

How do I summarise or classify a document with AI?

Read Document turns the file into text or Markdown; a Run Agent step does the rest. Put {{document.markdown}} in the agent's prompt, give the agent declared outputs such as risk and summary, and branch on them.

Where do the files come from and go to?

In: a form's Attachments field, a User Task, an HTTP download, storage connectors, a URL or inline content. Out: file objects saved to your configured storage, which the next step can email, upload, loop over, read again or hand to an AI agent.

Put documents in the loop.

Collect files with a form, read and compare them in the run, and send the result where it needs to go.

Document actions guide Forms and attachments