Performance and scaling

What takes time

Almost all time is spent waiting for the model provider. With the engine, a page takes 20 to 50 seconds (more for scans with tables and for handwriting); model only takes 8 to 28. Pages of one document are read in parallel, and so are the zoomed strips of a page.

Settings

Variable	Default	Effect
READANYDOC_WORKERS	4	Async jobs processed at once
READANYDOC_API_CONCURRENCY	32	Model requests in flight at once across the server. A rate-limit answer pauses all of them, then they resume
READANYDOC_DOCUMENT_TIMEOUT_SECONDS	1800	Longest time one document may take (504 timeout)

Sizing (starting points, not measured limits)

• Throughput is bounded by your provider's rate limits. Raise READANYDOC_API_CONCURRENCY only as far as your Anthropic or OpenAI tier allows.
• CPU is used by the layout model and OCR: 2 to 4 cores per instance is enough for the default settings.
• Memory: 2 to 4 GB per instance with the engine.
• Scale out by running more instances behind a load balancer with a shared volume for /data, or one instance per queue.

Cost control

• Use mode="model-only" (or READANYDOC_PROFILE=model-only) where layout and citations are not needed.
• For extract on long documents, settings.relevance_filter: true reads only the pages your schema needs.
• Every response reports usage with tokens and cost, so you can track spend per document.