ParseAI is Now Open Source

From today, the core document processing engine behind ParseAI is open source. You can read the code, run it locally, and see exactly how we turn a raw PDF or image into structured markdown or JSON.

The repository is at github.com/squrtech/parseai.

Why Open Source

Document AI is infrastructure. When you build a RAG pipeline, a loan processing system, or an insurance claims workflow on top of a document parsing API, you need to trust that the extraction is correct and understand why it fails when it does. That trust is hard to establish with a black box.

Open-sourcing the engine means you can read the parsing pipeline, understand the extraction logic, and reproduce results locally. You get the transparency of running your own system with the operational convenience of a managed API.

What's Inside

The engine is built on Docling, IBM Research's open-source document understanding library. Docling is not a PDF text extractor — it runs actual ML inference on every document:

• Layout analysis. IBM's Egret-Large vision model classifies every region on each page: heading, paragraph, table, figure, list, caption. This gives the parser a structural map before any text is extracted.
• Table structure recognition. IBM's TableFormer model reconstructs the row-column grid of detected tables — including merged cells, multi-row headers, and spanning columns.
• OCR. Tesseract handles scanned pages and image regions. Native-text PDFs are read directly; OCR is only applied where necessary.

On top of Docling, the ParseAI engine adds the infrastructure for running this at scale: a multi-tenant job queue, fair-use dispatcher (round-robin across organisations so one high-volume org cannot starve others), S3 result storage with presigned URLs, and the LLM extraction layer that maps parsed markdown to your schema with per-field confidence scores.

Supported Input Formats

The open-source engine accepts the same formats as the hosted API:

• PDF — native text and scanned
• JPEG, PNG, TIFF, BMP — document images
• DOCX, XLSX, PPTX — converted internally before layout analysis

Using the Hosted API vs Running It Yourself

You have two options now:

Hosted API. POST a document, get a job ID, poll for completion, download the result. No infrastructure to manage. We handle model loading, scaling, storage, and multi-tenancy. Per-page pricing — you pay only for what you process.

Self-hosted. Clone the repo, run docker compose up, and you have the full stack running locally: docling-server, Celery workers, Redis, MinIO, and the FastAPI backend. Good for air-gapped environments, compliance requirements, or teams that want to customise the pipeline.

Most teams find the hosted API faster to get to production. Self-hosting is there for when you need it.

Getting Started

If you want to use the hosted API, create an account at app.parseai.org, generate an API key under Settings → API Keys, and follow the quickstart guide.

If you want to run it yourself, star the repo, read the README, and run:

git clone https://github.com/ParseAI/parseai-parser
cd parseai-parser
cp .env.example .env
# fill in your API keys
docker compose up

Issues, PRs, and questions are welcome on GitHub. If you build something interesting on top of the engine, we would love to hear about it — reach us at sales@squrtech.com.