Most enterprise data stays locked in PDFs, scans, and slide decks. To use that data with LLMs or agents it must become structured JSON. Open‑source extraction lets you do this on‑premise, saving cost and keeping data private.
There are two distinct tasks hidden behind the phrase “PDF to JSON”. Schema‑driven extraction takes a document and a JSON schema and returns only the fields you need – ideal for invoices, forms, contracts. Document parsing rebuilds the whole page into JSON or Markdown, preserving layout, tables, reading order – useful for building RAG corpora.
For schema‑driven work consider lift (9 B vision model from Datalab) which uses schema‑constrained decoding to output valid JSON matching your schema, and NuExtract 3 (4 B from NuMind) which combines extraction with OCR and can also output Markdown. Both run locally via vLLM or Hugging Face and have CLI/Python APIs.
For document parsing look at Docling (IBM/LF AI) which outputs Markdown, JSON, DocTags and integrates with LangChain/LlamaIndex; the compact Granite‑Docling‑258M model does one‑shot conversion; MinerU (OpenDataLab) handles complex layouts and now uses a commercial‑friendly license; Marker (Datalab) adds an optional LLM flag for better tables; olmOCR 2 (Ai2) is a 7 B OCR‑focused model giving clean text/Markdown; DeepSeek‑OCR uses token compression for long documents; and Qwen3‑VL (Alibaba) serves as a general‑purpose VLM fallback when you need flexible prompting.
Licenses vary: most code is MIT or Apache‑2.0, while some model weights carry OpenRAIL‑M‑style restrictions; MinerU recently moved from AGPL‑3.0 to a custom Apache‑based license. Published benchmarks come from different test suites, so treat them as indicative and run your own samples before deciding.
Pick the tool that matches your task, check the license for your usage scale, and start with a small batch to validate accuracy and speed.
#AI #Product #OpenSource #LLM #DataExtraction #DocumentAI