thepipe

Extract clean data from tricky documents ⚡

thepi.pe is a package that can scrape clean markdown, multimodal media, and structured data from complex files. It uses vision-language models (VLMs) under the hood for superior output quality, and works out-of-the-box with any LLM, VLM, or vector database. It can extract well-formatted data from a wide range of sources, including PDFs, Word docs, Powerpoints, Python notebooks, videos, audio, and more. If you've tried PyMuPDF, Markitdown, Azure Document Intelligence, and Docling, this will be the last library you try.

Features 🌟

• Get clean markdown, tables, and images from any file
• Document AI that works out-of-the-box with any multimodal LLM
• AI-native file type detection, layout analysis, and structured data extraction

Get started in 5 minutes 🚀

Thepipe can be installed via the command line:

pip install thepipe-api

The default install only pulls in CPU-friendly dependencies so it is suitable for constrained environments and CI systems. GPU-enabled libraries such as PyTorch and Triton are left as optional extras.

Optional extras

The package exposes a set of extras so you can opt-in to heavier dependencies on demand:

Extra	Installs	When to use it
thepipe-api[audio]	openai-whisper	Local audio/video transcription via Whisper.
thepipe-api[semantic]	sentence-transformers	Semantic chunking with transformer embeddings.
thepipe-api[llama-index]	llama-index-core	Chunk.to_llamaindex() conversions.
thepipe-api[gpu]	PyTorch + Whisper + Sentence Transformers	Full GPU acceleration with VLM fine-tuning workloads.

If you are targeting CPU-only machines but still need the extras that depend on PyTorch, install the CPU wheels directly from the PyTorch index first and then add the extra. For example:

pip install torch --index-url https://download.pytorch.org/whl/cpu
pip install thepipe-api[semantic]

If you need full functionality with media-rich sources such as video and audio, you can choose to install the following system dependencies:

apt-get update && apt-get install -y ffmpeg

and use the global installation with pip:

pip install thepipe-api[all]

Default setup (OpenAI)

By default, thepipe uses the OpenAI API, so VLM features will work out-of-the-box provided you pass in an OpenAI client.

Custom VLM server setup (OpenRouter, OpenLLM, etc.)

If you wish to use a local vision-language model or a different cloud provider, you can provide a custom OpenAI client, for example, by setting the base url to https://openrouter.ai/api/v1 for OpenRouter, or http://localhost:3000/v1 for a local server such as OpenLLM. Note that uou must also pass the api key to your non-OpenAI cloud provider into the OpenAI client. The model name can be changed with the model parameter. By default, the model will be gpt-5.6-luna.

Scraping

Scraping clean markdown from a file is simple:

from thepipe.scraper import scrape_file

chunks = scrape_file("paper.pdf")

for c in chunks:
  print(c.text)

For enhanced scraping, you can pass in an OpenAI-compatible client and a vision-language model:

from openai import OpenAI
from thepipe.scraper import scrape_file

client = OpenAI()

chunks = scrape_file(
  filepath="paper.pdf",
  openai_client=client, # openai-like client; enables AI scraping
  model="gpt-5.6-luna" # vision-capable model
)

All available options are shown below:

from thepipe.scraper import scrape_file

chunks = scrape_file(
  filepath="paper.pdf", # path to the file to scrape
  verbose=True, # print progress logs
  chunking_method=chunk_by_page, # chunking function
  openai_client=client, # openai-like client; enables AI scraping
  model="gpt-5.6-luna", # vision-capable model (default: DEFAULT_AI_MODEL)
  include_input_images=True, # sends page images to VLM for enhanced scraping
  include_output_images=True, # include scraped images in the returned chunks
  max_input_image_size=1024, # max pixel size of the largest axis of images sent to the VLM (does not affect output images)
)

AI scraping works well when you need visual figures, bounding boxes, and OCR on image-only scans. Below, a skewed, stained scanned document with no text layer is scraped into clean markdown, with its tables and figures located:

Chunking

To satisfy token-limit constraints, the following chunking methods are available to split the content into smaller chunks.

• chunk_by_document: Returns one chunk with the entire content of the file.
• chunk_by_page: Returns one chunk for each page (for example: each PDF page or PowerPoint slide).
• chunk_by_length: Splits chunks by length.
• chunk_by_section: Splits chunks by markdown section.
• chunk_by_keyword: Splits chunks at keywords.
• chunk_semantic (experimental, requires sentence-transformers): Returns chunks split by spikes in semantic changes, with a configurable threshold.
• chunk_agentic (experimental, requires OpenAI): Returns chunks split by an LLM agent that attempts to find semantically meaningful sections.

For example,

from thepipe.scraper import scrape_file
from thepipe.chunker import chunk_by_document, chunk_by_page

# optionally, pass in chunking_method
# chunk_by_document returns one chunk for the entire document
chunks = scrape_file(
  filepath="paper.pdf",
  chunking_method=chunk_by_document
)

# you can also re-chunk later.
# chunk_by_page returns one chunk for each page (for example: each PDF page or PowerPoint slide).
chunks = chunk_by_page(chunks)

OpenAI Chat Integration 🤖

from openai import OpenAI
from thepipe.core import chunks_to_messages

# Initialize OpenAI client
client = OpenAI()

# Use OpenAI-formatted chat messages
messages = [{
  "role": "user",
  "content": [{
      "type": "text",
      "text": "What is the paper about?"
    }]
}]

# Simply add the scraped chunks to the messages
messages += chunks_to_messages(chunks)

# Call LLM
response = client.chat.completions.create(
    model="gpt-5.6-luna",
    messages=messages,
)

chunks_to_messages takes in an optional text_only parameter to only output text from the source document. This is useful for downstream use with LLMs that lack multimodal capabilities.

⚠️ It is important to be mindful of your model's token limit. Be sure your prompt is within the token limit of your model. You can use chunking to split your messages into smaller chunks.

LLamaIndex Integration 🦙

Install the optional extra and then call .to_llamaindex:

pip install thepipe-api[llama-index]

After installation, a chunk can be converted to LlamaIndex Document/ImageDocument with .to_llamaindex. Without the extra, a helpful error is raised instead of failing at import time.

Structured extraction 🗂️

Extraction uses OpenAI structured outputs, so results always conform to your pydantic schema.

from thepipe.extract import extract
from openai import OpenAI
from pydantic import BaseModel

client = OpenAI()

class LineItem(BaseModel):
    description: str
    amount_usd: float

results, tokens_used = extract(
    chunks=chunks,
    schema=LineItem,
    multiple_extractions=True,  # extract multiple rows of data per chunk
    openai_client=client
)

Running the test suite 🧪

Install the package in editable mode with the extras you rely on, then execute:

pip install -e ".[all]"
python -m unittest discover

Sponsors

Please consider supporting thepipe by becoming a sponsor. Your support helps me maintain and improve the project while helping the open-source community discover your work.

Visit Cal.com for an open-source scheduling tool that helps you book meetings with ease. It's the perfect solution for busy professionals who want to streamline their scheduling process.

Looking for enterprise-ready document processing and intelligent automation? Discover how Trellis AI can streamline your workflows and enhance productivity.

How it works 🛠️

thepipe uses a combination of computer-vision models and heuristics to scrape clean content from the source for downstream use with large language models, or vision-language models. You can feed the resulting markdown chunks directly into a model, or store them in a vector database such as ChromaDB, LLamaIndex, or an equivalent RAG framework.

Supported File Types 📚

Source	Input types	Multimodal	Notes
PDF	.pdf	✔️	Extracts page markdown and page images. With an OpenAI client, a VLM transcribes each page to markdown and returns cropped images/diagrams/charts instead of full page renders
Word Document	.docx	✔️	Extracts text, tables, and images
PowerPoint	.pptx	✔️	Extracts text and images from slides
Video	.mp4, .mov, .wmv	✔️	Uses Whisper for transcription and extracts frames
Audio	.mp3, .wav	✔️	Uses Whisper for transcription
Jupyter Notebook	.ipynb	✔️	Extracts markdown, code, outputs, and images
Spreadsheet	.csv, .xls, .xlsx	❌	Converts each row to JSON format, including row index for each
Plaintext	.txt, .md, .rtf, etc	❌	Simple text extraction
Image	.jpg, .jpeg, .png	✔️	Uses VLM for OCR in text-only mode
ZIP File	.zip	✔️	Extracts and processes contained files
Directory	any path/to/folder	✔️	Recursively processes all files in directory. Optionally use inclusion_pattern to pass regex strings for file inclusion rules.

Configuration & Environment

Set these environment variables to control API keys, hosting, and model defaults:

# Image hosting: instead of embedding base64 data URLs in LLM messages, save
# images to ./images and reference them as {HOST_URL}/images/{id}.jpg.
# HOST_URL must serve that folder (e.g. via your own API) and is required
# when HOST_IMAGES is true.
export HOST_IMAGES=true
export HOST_URL=https://your-domain.com

# Control scraping defaults
export DEFAULT_AI_MODEL=gpt-5.6-luna
export DEFAULT_EMBEDDING_MODEL=sentence-transformers/all-MiniLM-L6-v2

# Window length (in seconds) for splitting audio/video transcripts into chunks
export MAX_WHISPER_DURATION=600

Audio and video scraping (pip install thepipe-api[audio]) require the ffmpeg binary on your PATH.

CLI Usage

thepipe <source> [options]

AI scraping options

--openai-api-key=KEY To enable VLM scraping, pass in your OpenAI API key

--openai-model=MODEL Model to use for scraping (default is DEFAULT_AI_MODEL, currently gpt-5.6-luna)

--openai-base-url=URL Custom LLM endpoint, for local LLMs or hosted APIs like OpenRouter (default: https://api.openai.com/v1)

--ai_extraction ⚠️ DEPRECATED; will get API key from OPENAI_API_KEY environment variable

General scraping options

--text_only Output text only (suppress images)

--inclusion_pattern=REGEX Include only files whose _full path* matches REGEX (for dirs/zips)

--verbose Print detailed progress messages

Contributing

This package is quite opinionated in its design and implementation. Some modules are tightly coupled to the overall architecture, while others are designed to be hacked.

Before contributing, please create an issue on GitHub to discuss your ideas and how to best implement them. Pull requests that do not follow this process will be closed.