Stay organized with collections
Save and categorize content based on your preferences.
Custom extractor overview
Custom extractor extracts entities from documents of a particular type. For
example, it can extract the items in a menu or the name and contact information
from a resume.
Overview
The goal of the custom extractor is to enable Document AI users to build
custom entity extraction solutions for new document
types for which no pre-trained processors are available. Custom extractor includes
a combination of layout-aware deep learning models (for generative AI and custom
models) and template-based models.
Which training method should I use?
Custom extractor supports a wide range of use cases with three different modes.
Training method
Document examples
Document layout variation
Free form text or paragraphs
Number of training documents for production-ready quality, depending on variability
Fine tune and foundation model (generative AI).
Contract, terms of service, invoice, bank statement, bill of lading, payslips.
High to Low (preferred).
High.
Medium: 0-50+ documents.
Custom model.
Model.
Similar forms with layout variation across years or vendors (for example, W9).
Low to medium.
Low.
High: 10-100+ documents.
Template.
Tax forms with a fixed layout (for example, Forms 941 and 709).
None.
Low.
Low (3 documents).
Because foundation models typically require fewer training documents, they're
recommended as the first option for all variable layouts.
Confidence score
The confidence score communicates how strongly your model associates each entity
with the predicted value. The value is between zero and one, the closer it is to
one, the higher the model's confidence that the value corresponds to the entity.
This allows users to set triggers for manual review of individual entities when
the value is low. For example, determining whether the text in an entity is
"Hello, world!" or "HeIIo vvorld!"
The benefits of this approach allow for spotting individual entities with low
confidence, setting thresholds for which predictions are used, selecting the
optimal confidence threshold, and development
of new strategies for training models with higher accuracy and confidence scores.
For more information on evaluation concepts and metrics, see Evaluate
Performance
[[["Easy to understand","easyToUnderstand","thumb-up"],["Solved my problem","solvedMyProblem","thumb-up"],["Other","otherUp","thumb-up"]],[["Hard to understand","hardToUnderstand","thumb-down"],["Incorrect information or sample code","incorrectInformationOrSampleCode","thumb-down"],["Missing the information/samples I need","missingTheInformationSamplesINeed","thumb-down"],["Other","otherDown","thumb-down"]],["Last updated 2025-08-26 UTC."],[[["\u003cp\u003eCustom extractors are designed to identify and extract specific entities from various document types, including menus and resumes, for which pre-trained processors may not exist.\u003c/p\u003e\n"],["\u003cp\u003eThe custom extractor employs a combination of layout-aware deep learning models and template-based models to accommodate diverse document structures.\u003c/p\u003e\n"],["\u003cp\u003eThree training methods are available for the custom extractor: fine-tuning with foundation models, custom models, and template-based models, each suited for different levels of document layout variability.\u003c/p\u003e\n"],["\u003cp\u003eFoundation models are the preferred training option for documents with variable layouts, as they typically require fewer training documents compared to other methods.\u003c/p\u003e\n"],["\u003cp\u003eThe confidence score, ranging from zero to one, indicates the model's certainty in associating a value with a predicted entity, enabling users to set review thresholds and improve model accuracy.\u003c/p\u003e\n"]]],[],null,["# Custom extractor overview\n=========================\n\nCustom extractor extracts entities from documents of a particular type. For\nexample, it can extract the items in a menu or the name and contact information\nfrom a resume.\n\nOverview\n--------\n\nThe goal of the custom extractor is to enable Document AI users to build\ncustom entity extraction solutions for new document\ntypes for which no pre-trained processors are available. Custom extractor includes\na combination of layout-aware deep learning models (for generative AI and custom\nmodels) and template-based models.\n\nWhich training method should I use?\n-----------------------------------\n\nCustom extractor supports a wide range of use cases with three different modes.\n\nBecause foundation models typically require fewer training documents, they're\nrecommended as the first option for all variable layouts.\n\nConfidence score\n----------------\n\nThe confidence score communicates how strongly your model associates each entity\nwith the predicted value. The value is between zero and one, the closer it is to\none, the higher the model's confidence that the value corresponds to the entity.\nThis allows users to set triggers for manual review of individual entities when\nthe value is low. For example, determining whether the text in an entity is\n\"Hello, world!\" or \"HeIIo vvorld!\"\n\nThe benefits of this approach allow for spotting individual entities with low\nconfidence, setting thresholds for which predictions are used, selecting the\noptimal [confidence threshold](/document-ai/docs/evaluate#confidence_threshold), and development\nof new strategies for training models with higher accuracy and confidence scores.\n\nFor more information on evaluation concepts and metrics, see [Evaluate\nPerformance](/document-ai/docs/evaluate#all-labels)"]]