Open-Source PDF to JSON Extraction Models Emerge

In 2026, the challenge of converting enterprise data locked within PDFs, scans, and slide decks into structured JSON for large language models and agents is being addressed by a surge of open-source document extraction tools. These tools are crucial because LLMs and AI agents cannot process unstructured data until it is converted into a machine-readable format like JSON. The process of converting PDFs to JSON encompasses two distinct problems: schema-driven extraction, where predefined fields are populated with data, and document parsing, which reconstructs the document's layout and content into structured JSON or Markdown. Many organizations require one or both of these capabilities to effectively utilize their data.

Schema-driven extraction is particularly useful for documents with known fields, such as invoices, forms, contracts, and receipts. Users define a JSON schema, and the model populates it with corresponding values extracted from the document. This approach ensures that the extracted data conforms to a specific structure, making it readily usable for downstream applications. Document parsing, on the other hand, focuses on understanding the document's visual and textual structure. It identifies layout elements, reading order, tables, formulas, and code, then outputs this information as JSON or Markdown. This is essential for preparing clean datasets for retrieval-augmented generation (RAG) systems and AI agents.

The reliance on proprietary APIs for these tasks can be prohibitively expensive, with costs potentially reaching thousands of dollars per million pages, and raises privacy concerns due to the need to send sensitive documents off-premise. Open-source, open-weight models offer a compelling alternative by enabling local processing, thereby eliminating both the high costs and the privacy risks associated with cloud-based solutions. This shift towards local, open-source extraction is becoming the standard for many enterprises seeking cost-effective and secure data conversion.

Among the notable open-source models is Datalab's 'lift,' a 9B vision model designed for schema-driven extraction. Lift takes a JSON schema as input and guarantees valid JSON output through schema-constrained decoding. Built on the Qwen 3.5 architecture, it can be run locally via Hugging Face or on a remote vLLM server. Lift supports multi-page documents in a single processing pass, even for values that span across pages. The toolkit includes a command-line interface (CLI), a Python API, and a Streamlit-based 'Schema Studio' for schema creation and testing. The availability of such tools empowers organizations to maintain greater control over their data while unlocking its potential for AI applications.