PastaPress

PastaPress is a powerful command-line tool and Python module designed for stylistic text refinement via a local Ollama instance (e.g., Mac Studio). It pushes raw, messy text through an AI "press" and returns refined, smooth text without altering the core facts or structural integrity.

Read this documentation in German (Deutsch).

💡 Use Cases

• Free your text from unintended AI-typical markers — and declare the AI share yourself. Texts are increasingly prejudged by superficial "AI markers" (telltale phrasing patterns and artifacts) regardless of their actual substance. PastaPress rewrites text through a local press so it stands on its content again — the honest counterpart being that you declare the actual AI involvement openly, e.g. in the AI-disclosure statement of a scientific article. Transparency by declaration instead of prejudice by pattern.
• Polish raw notes into readable prose — meeting notes, drafts, and quick dumps come out fluent while every fact, list, and heading stays intact.
• Translate without breaking structure — Markdown, lists, and delimiters survive the round trip into any target language.
• Batch-press whole folders — queue a directory, let it run, keep the originals.

🌟 Features

• Chunk-based Processing: Processes text files paragraph by paragraph to bypass LLM context limits. Oversized paragraphs are split further at line and word boundaries.
• Flawless Reconstruction: Keeps delimiters, indentation, and original markdown formatting completely intact.
• Format Support: Supports .txt, .md, .json, .csv, .yaml, .tex, and auto-converts binary formats like .docx, .odt, and .rtf to clean Markdown using pypandoc. (Legacy binary .doc is not supported — convert it to .docx first.)
• Stylistic Control: Dynamically adapt the refinement style (gleichwertig, wissenschaftlich, einfach, kurz, or original).
• Translation Mode: Optionally translate text into any target language on-the-fly while preserving format.
• Queue System: Batch-process entire directories sequentially via queue.json.

🚀 Installation

Ensure you have Python 3.12+ installed.

git clone https://github.com/ellmos-ai/pasta-press.git
cd pasta-press
pip install -r requirements.txt

(Note: If you plan to process .docx or .odt files, the tool will attempt to download Pandoc automatically if it is missing.)

⚙️ Configuration

Configure your local Ollama host and default model (defaults to http://localhost:11434; settings are stored in a local, untracked config.json — see config.example.json):

python -m pastapress config --auto  # Auto-detects the best model on your host
# OR
python -m pastapress config --model qwen3.6:35b-mlx --host http://my-ollama-server:11434

Set your preferred default style and translation settings:

python -m pastapress config --style wissenschaftlich
python -m pastapress config --translate-mode on --lang "Spanish"

Model thinking/reasoning is disabled by default (~10-80x faster on thinking-capable models like qwen3.x; content stays intact, though small models may phrase things slightly rougher). Re-enable it if you prefer maximum polish over speed:

python -m pastapress config --thinking on

🛠️ Usage

Process a Single File

python -m pastapress process my_document.txt

Output will be saved as my_document_pasta-press.txt by default.

Override Styles and Languages per File

python -m pastapress process draft.docx --style original --translate English

Process a Directory (Batch / Queue)

python -m pastapress process ./my_folder
python -m pastapress process-queue

Process Raw Text (Integration)

python -m pastapress text "This is a very bad text that needs fixing."

🔒 Privacy & Data Security

• Local Processing: All data is processed completely locally via the configured Ollama host (default: http://localhost:11434).
• No Telemetry: No data is sent to external clouds or third-party APIs.
• Smart Filtering: (Planned - see ROADMAP.md) Future versions will offer strict tag/code filtering to prevent sensitive code chunks from being sent to the LLM.

📄 License & Provenance

MIT License — covers the code, prompts, and documentation in this repository (see LICENSE). Dependencies (requests, click, pypandoc) are installed via pip and keep their own licenses.

This tool was developed AI-assisted within the ellmos-ai ecosystem and is maintained with human review. See docs/ai-act-note.md for scope and intended use under the EU AI Act.