Extract text from PDF, Word (.docx), OpenDocument (.odt), RTF, and plain text files. Pure-Python where possible — no LibreOffice required. Single file or batch directory mode.

Document to Text

Extract plain text from PDF, Word, OpenDocument, RTF, and plain text files. Single file or batch directory mode. No LibreOffice required.

Why this exists

Most document-analysis jobs (resume screening, contract review, expense extraction) need the text content of a file — not its visual rendering. Going through PDF is expensive: it requires LibreOffice (~700 MB) for .doc/.docx/.odt conversion, and the AI still has to re-extract text from the rendered PDF.

This tool extracts text directly from the source, with zero install friction for modern formats (.txt, .docx, .odt — pure stdlib). PDF and RTF need small pip packages; the script prints a clear install command if a backend is missing.

For workflows that need rendered-PDF fidelity (form-field positions, images, page-perfect layout), use a doc-to-pdf converter instead.

Inputs

Property	Required	Default	Description
`inputFile`	one of two	—	Path to a single document.
`inputDir`	one of two	—	Folder whose matching files are all processed.
`extensions`	—	`pdf,docx,odt,rtf,txt`	Comma-separated extensions when using `inputDir`.
`outputFile`	—	—	If set, all extracted text is concatenated and written here with per-file headers.
`outputDir`	—	—	If set, writes one `.txt` per input document.
`maxBytes`	—	`0`	Per-file byte cap (`0` = unlimited).

Provide either inputFile or inputDir, not both.

Outputs (published)

Name	Description
`text`	Full concatenated text with per-file headers.
`fileCount`	How many input files matched.
`successCount`	How many extracted cleanly.
`failureCount`	How many errored.
`outputFiles`	Comma-separated list of files this task wrote.
`errors`	Newline-separated error messages, one per failed file.

Format support

Extension	Backend	Install required?
`.txt`	stdlib (utf-8, utf-16, latin-1 fallback)	None
`.docx`	stdlib (`zipfile` + `xml.etree`)	None
`.odt`	stdlib (`zipfile` + `xml.etree`)	None
`.rtf`	`striprtf` (pip) — regex fallback if missing	`pip install striprtf` (optional)
`.pdf`	`pypdf` or `pdfminer.six`	`pip install pypdf`
`.doc`	`antiword` or `catdoc` binary	`apt install antiword catdoc` (or `brew install antiword`)

The legacy binary .doc format is rare in practice — most modern Word files are .docx. If you only have .doc originals, install antiword or catdoc (the script tries both), or save them as .docx first.

Installing the pip packages on Debian/Ubuntu/Mint (PEP 668)

Modern Debian-based distros mark the system Python as "externally managed" and refuse pip install without an extra flag. Three options, in order of cleanliness:

Use the runtime env's Python directly (recommended for InTouch): pip install --user --break-system-packages pypdf striprtf
Set up a virtualenv and point your python runtime env at it (no global state changes).
Skip PDFs/RTFs — the .docx/.odt/.txt paths work with zero dependencies. The tool still succeeds on those and reports per-file errors for the formats whose backend is missing.

Chaining patterns

Resume folder → Claude assessment — document-to-text on /path/to/resumes → publishes text → feed to anthropic task as {{document-to-text.text}} in the prompt.
Invoice email attachments → SQL — pull attachments, run them through document-to-text, then parse and INSERT.
Knowledge-base ingest → embeddings — extract text from a doc folder, chunk with document_chunker, embed with embeddings.

Example

name: resume-review
tasks:
  - name: extract
    tool: document-to-text
    properties:
      inputDir: /home/me/resumes
      maxBytes: "100000"

  - name: assess
    tool: anthropic
    credential: claude-api
    properties:
      prompt: |
        Assess the following resumes. For each candidate, extract skills,
        experience level, and a 1-10 rating.

        {{extract.text}}
      maxTokens: 4000
      outputFile: /home/me/resume-assessment.txt