Skip to content

Requirements & sizing

The app itself

The backend is I/O-bound: it waits on the LLM and OCR endpoints and does very little compute of its own. Nothing is persisted, so there is no storage to size.

Minimum Comfortable
CPU 2 cores 4 cores
RAM 2 GB 4 GB
Disk ~2 GB for the images
GPU none only for a co-located OCR/LLM sidecar

RAM scales with concurrent documents, not with corpus size. The dominant consumers are the rendered page images of a scanned PDF during OCR and the detection cache (bounded at 100 entries / 15 minutes).

Supported platforms: Linux with Docker or Podman. macOS works for development.

What you actually need to size

The models, not the app:

  • Detection LLM — an OpenAI-compatible endpoint. Throughput here sets your throughput: a document is split into chunks, and every chunk runs LLM_DETECTION_PASSES times (default 2), plus one re-check call per document. LLM_MAX_CONCURRENT_REQUESTS (default 4) caps the total in flight across all documents.
  • OCR — only if you process scans. A vision-LLM engine needs a GPU; docling-serve with Tesseract runs on CPU. VISION_OCR_MAX_CONCURRENT_PAGES (default 2) caps page requests globally.

See LLM endpoints and OCR engines.

Network

  • Users → frontend (port 8080 by default), behind your own auth proxy.
  • backend → the configured LLM/OCR endpoints. Nothing else.
  • No outbound internet access is required at runtime. There is no telemetry, no CDN, and no third-party font or script.

The backend container publishes no port at all: nginx in the frontend container proxies /api/ to it over the compose network.

Browsers

Any current Chromium, Firefox, or Safari. The redacted-PDF preview uses the browser's built-in PDF viewer; where that is unavailable the download still works.

Before you process real data

Technical requirements are the easy part. Also settle:

  • Who the data controller is and which legal basis applies (DPIA template).
  • Where document content is allowed to flow (Data flow).
  • Who reviews the output, and against what standard (Security overview).
  • How well the pipeline performs on your document types (Evaluation).