← All models

Document AI Custom Extractor v1.4

Published Publisher Googlepretrained-foundation-model-v1.4-2025-02-05

Model Overview

A version of Google Cloud Document AI's custom document information extraction model, powered by Gemini 2.0 Flash. Users can define the fields to extract and convert document content into structured entities. (docs.cloud.google.com)

Capabilities and Uses

  • Extracts document fields according to a custom schema, suitable for processing business documents in different formats and layouts. Custom Extractor offers zero-shot, few-shot, and fine-tuning approaches; zero-shot requires only a schema definition, while fine-tuning can use labeled documents to further improve performance on specific tasks. (docs.cloud.google.com)
  • v1.4 supports confidence scores. On August 29, 2025, Google announced that signature detection and derived entity detection for this version had reached general availability; the latter can be used to infer fields not directly stated, such as inferring a country from an address or counting items in a table. (docs.cloud.google.com)

Release and Status

  • February 14, 2025: Released in public preview in the US and EU regions; the announcement reported improved accuracy. (docs.cloud.google.com)
  • March 19, 2025: The model became generally available (GA); fine-tuning in the US and EU regions was in preview at the time. (docs.cloud.google.com)
  • Deprecation: The official deprecation list gives March 31, 2026 as the deprecation date for this version. Actual availability and migration arrangements should be determined by the project console and the latest Google Cloud notices; a specific shutdown time cannot be asserted from the deprecation date alone. (docs.cloud.google.com)

Limitations

  • The initial preview announcement listed an online processing quota of up to 120 pages per minute; this was the quota at release and is not guaranteed to apply to current projects. Starting with v1.4, online quotas are measured in pages by processor type and foundation model version. (docs.cloud.google.com)
  • The general per-request limits listed in the current Custom Extractor documentation are: 15 pages for online synchronous processing, 200 pages for offline batch processing, and 30 pages for online synchronous processing in imageless mode; documents containing checkboxes are limited to 10 pages for both online and batch processing. Specific versions and project configurations may have additional limits. (docs.cloud.google.com)
  • Extraction results may be incorrect. Test sets, confidence scores, and human review should be used together to determine whether results meet business requirements; field naming also affects extraction performance. (docs.cloud.google.com)

Pricing

Query date: November 19, 2026. Google's official standard pay-as-you-go prices for Custom Extractor are in US dollars (USD): $30 per 1,000 pages for the first 1 million input pages per month and $20 per 1,000 pages for pages exceeding 1 million. Google states that generative AI processor versions have the same extraction call charges as custom models; billing here is based on the number of input pages processed, with no separate output token price listed. Failed requests (4xx or 5xx) are not charged. This product-level pricing does not mean that the deprecated v1.4 is still available for use. (cloud.google.com)

Google also lists the Gemini Enterprise Flexible Savings Plan, which requires eligibility and may require enrollment: the one-year prices are $27 and $18 per 1,000 pages, respectively, and the three-year prices are $24 and $16; the two prices correspond respectively to the first 1 million pages per month and the pages beyond that. Separate pricing for fine-tuning training has not been verified, and other related Google Cloud services may incur additional charges. (cloud.google.com)

Official Sources