The distinction that matters is between converting a document to Markdown, which gives a model everything, and extraction, which takes a key or a table_header and returns only that — the difference between putting a page in the context window and putting a field there. If you refer to the extraction tool by name in a prompt, note that it is spelled general_information_extration.
A document-recognition server over TextIn, three tools that each take a local path: text recognition, targeted extraction, and Markdown conversion.
- Text recognised out of images, Word documents and PDF files
- Named values pulled from a document by key, and tables pulled out by their header row
- The same three formats converted to Markdown, structure intact, for a model to read
APP_ID and APP_SECRET for a TextIn account, with MCP_SERVER_REQUEST_TIMEOUT alongside them in the same configuration. Every tool takes a path, so the documents have to sit where the server can open them.
One command plus a key — npx -y @intsig/server-textin, then supply credentials