ParseDoc

Document → AI → Markdown. A local-first CLI that converts documents into clean, structured, AI-ready Markdown.

ParseDoc turns PDFs, Word docs, PowerPoint decks, HTML pages, plain text, and images into well-structured Markdown. Libraries extract the raw facts, an optional AI model understands the structure, and a deterministic renderer produces the final output. It works fully offline (local rule-based structuring) and can optionally use an AI provider for smarter structuring.

Features

• Local-first: conversion works with no AI provider configured (deterministic fallback).
• Multi-format input: PDF, DOCX, PPTX, HTML, TXT/MD, and images (PNG/JPG) via OCR.
• AI structuring (optional): local, ai, or hybrid modes.
• Pluggable AI providers: OpenAI-compatible (vLLM/LM Studio), OpenAI, Google Gemini, and Ollama (native, no SDK needed).
• Image extraction: pull embedded images from DOCX into an assets folder.
• OCR: Tesseract-backed text extraction for scanned PDFs and images.
• Multiple output formats: markdown (default), json, html, text.

Installation

git clone <repo-url>
cd parsedoc
pip install -e .

This installs the parsedoc command. (Requires Python 3.10+.)

Optional dependencies

CLI Usage

The CLI is built with Typer. Run parsedoc --help for the full list.

Commands

convert

parsedoc convert INPUT_FILE \
  --output out.md \
  --format markdown \
  --mode hybrid \
  --ai-provider ollama \
  --model qwen3 \
  --extract-images

Options

batch

parsedoc batch ./docs --pattern "*.docx" --format markdown --extract-images

inspect

parsedoc inspect INPUT_FILE --format json     # full extracted structure
parsedoc inspect INPUT_FILE --format summary  # blocks / title / format

Configuration

ParseDoc stores config as TOML (default location: ~/.config/parsedoc/config.toml, or via default_config_path()). You can edit it directly or use the CLI.

parsedoc config list
parsedoc config set --key ai_provider --value ollama
parsedoc config set --key model --value qwen3
parsedoc config reset

Environment variables

These override the TOML/config values at runtime:

Integrating AI Providers

ParseDoc supports four provider types, selected via --ai-provider (CLI), the ai.provider config key, or PARSEDOC_AI_PROVIDER.

Provider is chosen by the factory in parsedoc/ai/base.py:build_provider.

1. OpenAI-compatible (default)

This is the default and works with most self-hosted / drop-in OpenAI servers.

export PARSEDOC_BASE_URL="http://localhost:11434/v1"   # e.g. Ollama's OpenAI shim
export PARSEDOC_API_KEY="local"                         # or your real key
export PARSEDOC_MODEL="qwen3"
parsedoc config set --key ai_provider --value openai-compatible

TOML equivalent (~/.config/parsedoc/config.toml):

[ai]
enabled = true
provider = "openai-compatible"
base_url = "http://localhost:11434/v1"
model = "qwen3"
api_key = "local"
temperature = 0.2
max_tokens = 4096

2. OpenAI (hosted)

export PARSEDOC_BASE_URL="https://api.openai.com/v1"
export PARSEDOC_API_KEY="sk-..."
export PARSEDOC_MODEL="gpt-4o-mini"
parsedoc config set --key ai_provider --value openai

3. Google Gemini

export PARSEDOC_API_KEY="AIza..."
export PARSEDOC_MODEL="gemini-1.5-flash"
parsedoc config set --key ai_provider --value gemini

Note: the Gemini provider uses its own endpoint; base_url is optional and falls back to the Google Generative Language API.

4. Ollama (native)

No Python SDK required — ParseDoc talks to Ollama over HTTP using the standard library.

export PARSEDOC_BASE_URL="http://localhost:11434"   # Ollama root, not the /v1 shim
export PARSEDOC_MODEL="qwen3"
parsedoc config set --key ai_provider --value ollama

Then make sure the model is pulled:

ollama pull qwen3

Processing Modes

Set via --mode on convert/batch, or output/ai config sections (the provider/enabled keys).

Supported Input Formats

Check support before converting:

python -c "from parsedoc.core.detection import is_supported; print(is_supported('file.docx'))"

Using ParseDoc as a Library

from parsedoc.core.config import Config
from parsedoc.core.pipeline import Pipeline

config = Config().load_from_file()
pipeline = Pipeline(config)

markdown = pipeline.process(
    "report.docx",
    output_format="markdown",
    mode="hybrid",
    extract_images=True,
    output_path="report.md",
)

License

See repository for license details.