<2-3 sentences summarising what the project addresses and solves, the target audience, and the benefit.> Datalab-to/lift offers a 9B vision model for extracting structured JSON from PDFs and images, addressing document intelligence needs. It targets engineers and power users by providing schema-constrained decoding, ensuring valid JSON output that precisely matches user-defined schemas for streamlined data processing.

How It Works

Lift utilizes a 9B vision model with schema-constrained decoding to guarantee JSON output adheres strictly to provided schemas. It handles multi-page documents holistically, even across page breaks. Inference options include local execution via HuggingFace Transformers or optimized remote serving using a vLLM server.

Quick Start & Requirements

Installation is via pip: pip install lift-pdf (base/vLLM) or pip install lift-pdf[hf] (HuggingFace, requires torch). The vLLM server can be launched with lift_vllm. A Schema Studio app for schema development is available via pip install lift-pdf[app].

Highlighted Details

• Schema-Constrained Decoding: Ensures valid JSON output matching user schemas.
• Multi-Page Handling: Processes entire documents, including cross-page data, in a single pass.
• Flexible Inference: Supports local HuggingFace and optimized remote vLLM serving.
• Schema Studio: Interactive app for schema creation and testing.
• Performance: Achieves 90.2% field accuracy and 9.5s median latency on benchmarks, competitive for local inference.

Maintenance & Community

The project relies on Huggingface Transformers and vLLM. No specific community channels or contributor details are provided in the README.

Licensing & Compatibility

Code is Apache 2.0. Model weights use a modified OpenRAIL-M license: free for research, personal use, and startups (<$5M revenue). Competitive use against Datalab's API is prohibited. Broader commercial licensing, including on-premise, requires contacting the provider.

Limitations & Caveats

The OpenRAIL-M license restricts commercial use, especially for competitive offerings or on-premise deployments. The Datalab API offers higher accuracy (95.9% field accuracy) and features like per-field verification and citations. Schema definition is limited, excluding complex types like enum or anyOf.