Limits

This document lists the system limits that apply to Document AI. Unlike quotas, system limits can't be changed.

Content limits

The following content limits apply to all Document AI processors.

Content limit	Value
Maximum image resolution (limit does not apply to PDF files)	40 megapixels (per page if image contains multiple pages)
Maximum file size for online processing requests	20 MB
Maximum file size for batch processing requests	1 GB
Files per batch processing request	5,000 files
Human-in-the-Loop pages per document	10 pages

Processor limits

Limits are defined in the current list.

Extraction processors

Classification processors

Digitize processors

Pretrained processors

Limitations for Document AI

Document AI has the current limitations.

Criteria	Stable release July 2023
Dataset	Maximum of 30,000 documents total Maximum of 250,000 pages total
Document import	Maximum of 5,000 documents per import Maximum of 200 pages per document
Limits to train a Custom Document Extractor (CDE)	Model-based training (GA) Training dataset maximums: 25,000 documents; 100,000 pages Training dataset minimum: each label needs to be present on at least 1 label per 10 documents Test dataset maximums: 2,000 documents; 8,000 pages Test dataset minimum: every label on at least 10 documents Maximum of 200 pages per document Template-based training (GA) Training dataset maximums: 300 documents, 300 pages Training dataset minimum: every label on at least on at least 3 documents Test dataset maximums: 2,000 documents; 8,000 pages Test dataset minimum: every label on at least 3 documents Maximum of 20 pages per document
Limits to train a Custom Document Classifier (CDC) or a Custom Document Splitter (CDS)	Training dataset maximums: 30,000 documents; 100,000 pages Training dataset minimum: every label on at least 10 documents Test dataset maximums: 2,000 documents; 8,000 pages Test dataset minimum: every label on at least 2 documents Maximum of 200 pages per document
Labeling	To get started, ensure document labels meet defined minimum training and evaluation thresholds. To begin evaluating model performance for documents with layout variation, label at least 100 documents. Specifically, ensure that each label exists on 50 documents in training and 50 in evaluation. Maximum allowed labels (fields): 150 Label size limits (characters): Long items aren't well supported, but there's no explicit limit. Chunk documents into 800- or 1,000-token pieces, with 100 to 200 tokens overlapping between chunks. (Items longer than the overlapping area might run into quality issues.) Label occurrences in a document: No limit
Geographic coverage	Regions generally supported: US, EU (multiregion) Regions with limited accessibility: Germany, Singapore, UK, Canada, India, Australia