Skip to main content
pdf-online completes the “parse a file → publish to a workspace” workflow inside an agent conversation. Give the agent a file, destination platform, and target type. It uses SoMark to parse the content, then creates editable content in Feishu, DingTalk, or Notion. Use it to turn scans, contracts, papers, textbooks, or reports into online documents; convert complex tables into spreadsheets; or import lists, plans, and progress trackers as filterable records. When you publish to multiple platforms in one task, the source is parsed only once.
PDF Online creates editable content using the destination platform’s native capabilities. You can continue editing after publishing, and a routine check of key data is recommended.

​
What you can create

PDF Online can publish the same parsed result as a document, spreadsheet, or structured-record target. Choose the content form based on how you will work with it, then select the destination platform: If you omit the target type, Feishu defaults to a cloud document and Notion defaults to a page. DingTalk selects a route based on narrative, grid, or record intent; the agent asks when the choice would materially change fidelity. A table in the source never causes Notion to switch to database mode automatically. After you select a target, the Skill maps the structures recognized by SoMark into each platform’s native content model rather than pasting plain text:

​
Feishu

Cloud document supports headings, body text, bulleted and numbered lists, quotes, to-dos, code, tables, images, and inline and block equations. Choices and footnotes are adapted as lists and quotes, respectively. Spreadsheet supports worksheets, text cells, merged cells, borders, cell styles such as font, alignment, and wrapping, row and column dimensions, and embedded images. Bitable supports text, number, and date fields with their records.

​
DingTalk

Online document supports headings, body text, lists, code, images, tables, and equations; table cells can use rich text, images, equations, and superscripts. Choices are adapted as lists and footnotes remain as numbered, editable endnotes. Spreadsheet supports worksheets, text cells, merged cells, styles, row and column dimensions, and images. AI Table supports text, date, progress, number (including currency and rating), and single-select fields with their records; people fields are created only when valid userId and corpId values are provided.

​
Notion

Page supports a page title, headings, body text, lists, a table of contents, tables, images, code blocks, and equation blocks. Choices are adapted as bulleted lists and footnotes as quotes; stamps and chemical structures become images, table-cell images become links followed by separate images, and the API cannot create merged cells. Database supports title, rich-text, number, date, and single-select fields with their records.

​
How it works

The same source is parsed at most once per task. Multiple platforms share the same Markdown and JSON pair. If publishing fails, recovery reuses those results instead of consuming parsing quota again. The Skill skips parsing only when you explicitly provide an exact matching SoMark Markdown and JSON pair. It never searches for adjacent, same-name, or historical results. If you provide both a source file and the pair, Markdown and JSON are authoritative for content and structure; the source file is provenance only.

​
Before you begin

1

Install both Skills

pdf-online selects and publishes the target. For raw files, it calls the separately installed official somark-document-parser Skill.
2

Configure your SoMark API key

For the overseas service, get an API key at somark.ai. For the Mainland China service, use somark.cn. Configure SOMARK_API_KEY as instructed by your agent.SoMark parsing consumes quota. Reusing an explicitly named result pair does not parse the file again.
3

Install platform dependencies and authorize access

All three adapters require a working Python 3 environment. Prepare the dependency for your destination:Notion also requires a parent-page link where the connection can create content.
4

Prepare the input and target

Raw input can be a PDF, common image format, Word document, or PowerPoint presentation. Prepare the file path, platform, target type, and an optional destination link. Never put API keys, access tokens, or company credentials in a prompt.

​
Basic usage

The display name is PDF Online, and the Skill name is pdf-online. Include $pdf-online in your prompt and name the input, platform, and target type for more reliable routing.

​
Publish to multiple platforms

​
Limitations

  • The Skill does not rewrite OCR prose, invent missing content, or automatically repair vertical text, alignment, or damaged formulas.
  • Table conversion follows the structure recognized by SoMark. It does not run a second OCR pass or visually infer headers.
  • Spreadsheet images, merge ranges, and styles are processed independently; an issue with one element does not affect content already written to the workbook.
  • Structured-record targets need clear headers. Empty or duplicate headers receive safe names and a mapping notice; the agent never guesses among multiple tables.
  • After a creation or post-processing failure, the Skill reuses the current parse and existing target instead of parsing again or creating an arbitrary replacement.
  • In a multi-platform task, each destination is independent. Failure on one platform does not undo successful publishing on another.