datalab-to/lift

Extract structured data from documents quickly and accurately.

What it solves

lift is designed to extract structured JSON data from PDFs and images. It solves the problem of converting unstructured document intelligence into machine-readable formats by allowing users to define a specific JSON schema that the model must follow.

How it works

It utilizes a 9B vision model that processes documents (including multi-page PDFs) in a single pass. The system employs schema-constrained decoding to ensure that the output is a valid JSON object that strictly matches the provided schema. Users can run the model locally via HuggingFace or through a vLLM server for optimized production and batch processing.

Who it’s for

This tool is for developers and data engineers who need to automate the extraction of specific fields from documents like invoices or reports, and those who require guaranteed JSON output for downstream applications.

Highlights

  • Schema-Constrained Decoding: Guarantees the output is valid JSON matching the user's schema.
  • Multi-page Support: Handles documents where values may span across multiple pages in a single pass.
  • Flexible Inference: Supports both local HuggingFace execution and remote vLLM server deployments.
  • Schema Studio: Includes a Streamlit-based interactive app for building and testing schemas.
  • High Performance: Benchmarked with high field accuracy and lower latency compared to several other models.

Related

  • Project
    GoogleCloudPlatform/genmedia-creative-studio

    A web application for creative exploration that integrates Google Cloud's generative media models for image, video, music, and speech generation.

  • Project
    lucidrains/perceiver-pytorch

    A PyTorch implementation of the Perceiver and Perceiver IO architectures, designed to process diverse data types like images and sequences through a unified, efficient iterative attention mechanism.

  • Project
    zju3dv/MatchAnything

    MatchAnything is a research project that trains a single neural network to find correspondences between images from any visual modality (e.g., RGB, infrared, sketches). Pre‑trained weights are hosted on Hugging Face, and an online demo lets users try cross‑modality matching instantly. The approach aims to replace many specialized matching pipelines with one universal model, useful for robotics, remote sensing, medical imaging, and AR.

  • Project
    lxtGH/OMG-Seg

    A unified framework for visual perception and reasoning that combines image, video, and pixel-level segmentation into a single model, reducing the need for specialized specialist models.