Unstructured Inference

Warnings

• gotcha Detectron2 is a crucial dependency for using many layout parsing models within unstructured-inference, particularly those from the layoutparser model zoo. It is NOT automatically installed with `pip install unstructured-inference` and its installation can be complex, especially on Windows, where it's not officially supported. Users on macOS/Linux may need to build it from source.
  Fix: Refer to the Unstructured-IO documentation or Detectron2's installation guide for specific instructions. For macOS/Linux, `pip install 'git+https://github.com/facebookresearch/detectron2.git@57bdb21249d5418c130d54e2ebdc94dda7a4c01a'` is often required. Windows users may need to find community-supported workarounds or use WSL.
• breaking The library has a strict Python version requirement, currently supporting Python 3.12 only. Using other Python versions will lead to installation or runtime errors.
  Fix: Ensure your environment is running Python 3.12. Consider using virtual environments (like `venv` or `uv`) to manage specific Python versions for your projects.
• gotcha When `unstructured-inference` is used in conjunction with the main `unstructured` library, it's crucial to keep both packages synchronized to avoid unexpected behavior or errors, as `unstructured-inference` provides the underlying model capabilities for `unstructured`'s partitioning bricks.
  Fix: When updating `unstructured`, ensure `unstructured-inference` is also updated to a compatible version. Using `pip install "unstructured[all-docs]"` (if using `unstructured`) often helps ensure dependent packages are aligned.
• gotcha There have been reports of issues with table extraction functionality in recent versions of `unstructured-inference`, where the latest versions may not extract tables as effectively as older versions.
  Fix: If experiencing issues with table extraction, consult the GitHub issues for `unstructured-inference` for potential workarounds or targeted fixes. Downgrading to a previous version known to work with table extraction might be a temporary solution, but use with caution regarding other potential regressions.
• deprecated When using `unstructured`'s `partition` function with `strategy='hi_res'` (which utilizes `unstructured-inference` models), the `model_name` parameter is deprecated. Users should now use `hi_res_model_name` instead.
  Fix: Update calls from `partition(..., model_name='yolox')` to `partition(..., hi_res_model_name='yolox')`.

Install

• pip install unstructured-inference Basic Installation
• pip install 'git+https://github.com/facebookresearch/detectron2.git@57bdb21249d5418c130d54e2ebdc94dda7a4c01a' Detectron2 (for layoutparser models on Linux/macOS)

Imports

• DocumentLayout

  from unstructured_inference.inference.layout import DocumentLayout

• get_model

  from unstructured_inference.models.base import get_model

Quickstart

This quickstart demonstrates how to load a PDF document and extract its layout elements using the default inference model. It creates a dummy PDF for immediate execution. In a real application, you would replace `temp_pdf_path` with the path to your actual PDF file. The output includes the detected element types and their truncated text content.

import os
import tempfile

# Create a dummy PDF file for demonstration
# In a real scenario, you would provide the path to your actual PDF.
with tempfile.NamedTemporaryFile(suffix=".pdf", delete=False) as temp_pdf:
    temp_pdf_path = temp_pdf.name
    temp_pdf.write(b"%PDF-1.4\n1 0 obj<</Type/Catalog/Pages 2 0 R>>endobj 2 0 obj<</Type/Pages/Count 1/Kids[3 0 R]>>endobj 3 0 obj<</Type/Page/Parent 2 0 R/MediaBox[0 0 612 792]/Contents 4 0 R>>endobj 4 0 obj<</Length 41>>stream\nBT /F1 24 Tf 100 700 Td (Hello Unstructured!) Tj ET\nendstream\nendobj\nxref\n0 5\n0000000000 65535 f\n0000000009 00000 n\n0000000055 00000 n\n0000000108 00000 n\n0000000201 00000 n\ntrailer<</Size 5/Root 1 0 R>>startxref\n294\n%%EOF")

from unstructured_inference.inference.layout import DocumentLayout

try:
    # Perform layout parsing on the document
    # For real use, replace temp_pdf_path with your PDF file path.
    layout = DocumentLayout.from_file(temp_pdf_path)

    print(f"Found {len(layout.pages)} page(s) in the document.")
    for i, page in enumerate(layout.pages):
        print(f"--- Page {i+1} ---")
        for element in page.elements:
            print(f"Element Type: {element.type}, Text: {element.text[:50]}...")
            # You can also access bounding box, model name, etc.
            # print(f"  Bounding Box: {element.bbox}, Model: {element.detectron_model_name}")
finally:
    # Clean up the dummy PDF file
    os.remove(temp_pdf_path)