Amazon Textract features
Why Amazon Textract?
Amazon Textract is a machine learning (ML service) that automatically extracts text, handwriting, layout elements, and data from documents and images. It goes beyond simple optical character recognition (OCR) to turn document data into machine readable text. Amazon Textract Insights brings document understanding to Amazon Textract with no model training required. Describe the information needed in natural language and receive structured data from multi-document packets. Each answer returns a confidence score and page-level grounding so you can make informed decisions about how to use the results.
Amazon Textract Core
Open allAmazon Textract Core automatically detects printed text in 40+ languages, including more than 30 European languages, plus Arabic, Chinese (Simplified and Traditional), Hebrew, Hindi, Japanese, and Korean. It handles right-to-left and mixed-script documents and also recognizes English handwriting. Its ML-powered OCR reads text in a wide range of fonts and styles, including long words and noisy or distorted text. Every word and line comes with a bounding box and a confidence score.
Amazon Textract Core preserves the structure of tables during extraction, including rows, columns, and merged cells, in every supported language. It handles skewed or rotated tables, too. This helps with documents made up largely of structured data, such as financial reports or medical records. You can automatically load the extracted data into a database using a predefined schema. For example, rows of item numbers and quantities in an inventory report keep their association, so an inventory management application can easily update item totals.
Amazon Textract Core detects key-value pairs in documents automatically and keeps their context without manual work. A key-value pair is a set of linked data items. For example, in a document, the field "First Name" is the key, and "Jane" is the value. Amazon Textract Core also detects checkboxes and other selection elements. This makes it easy to import the extracted data into a database or use it as a variable in an application. With traditional OCR solutions, keys and values are extracted as natural text, and their relationship is lost unless you write and maintain hard-coded rules for each form. Forms extraction supports English, French, German, Italian, Portuguese, and Spanish documents.
Ask for the data you need as natural language questions, such as "What is the customer's name?", and get the exact answer, such as "John Doe", in the API response. You don't need to know how the data is structured in the document (table, form, implied field, or nested data), or handle variations across document versions and formats. Queries are pretrained on a wide range of documents, including pay stubs, bank statements, W-2s, loan application forms, mortgage notes, claims documents, and insurance cards. This reduces the need for post-processing, manual review of extracted data, or training ML models. Queries supports English documents.
Customize the pretrained Queries feature to improve extraction accuracy on document types specific to your business, while you keep control and ownership of your data. In the AWS Management Console, you can upload as few as 10 sample documents, annotate the data, and train a custom adapter within a few hours. Custom Queries supports English documents.
Extract layout elements such as titles, headers, section headings, paragraphs, lists, figures, page numbers, and footers, and get the content in its natural reading order, even on complex multi-column pages. Layout is a feature type in the Analyze Document API. You can use it on its own or with other Analyze Document feature types.
Detect signatures on any document or image. This makes it easy to find signatures on documents such as checks, loan application forms, and claims forms. The API response includes the location of each signature and its confidence score.
Invoices and receipts come in a wide variety of layouts, which makes extracting their data by hand at scale slow and difficult. Amazon Textract Core uses ML to understand the context of invoices and receipts. It automatically extracts data such as vendor name, invoice number, item prices, total amount, and payment terms, including data that isn't explicitly labeled, such as a vendor name that appears only in a logo or line items without column headers. It returns normalized key names, so you can easily compare output across many documents. Analyze Expense supports English documents.
Amazon Textract Core uses ML to understand the context of identity documents such as U.S. passports and driver's licenses, without templates or additional configuration. You can automatically extract specific information, such as date of expiry and date of birth, and identify implied information, such as name and address. Businesses that provide ID verification and those in finance, healthcare, and insurance can use Analyze ID to automate account creation, appointment scheduling, employment applications, and more from a picture or scan of an identity document.
Analyze Lending is a managed, preconfigured intelligent document processing API that automates the extraction of information from mortgage loan packages. Upload your loan documents, and its prebuilt ML models classify and split the package by document type, then route each page to the right analysis to extract its data. Analyze Lending supports English documents.
To process document types beyond Analyze Lending's predefined list, or to define your own fields, use document splitting and blueprint matching in Amazon Textract Insights.
Amazon Textract Insights
Open allDescribe the fields you need, and get them back as structured data, named the way you defined them. A blueprint is a schema that lists the fields to extract from one type of document. Each field has a name, a data type, and a natural language description of what to find and how to format it. Start with one of more than 35 public blueprints that AWS maintains as common documents, such as Form W-2s, Form 1040s, invoices, bank statements, pay stubs, passports, and driver's licenses. You can also create your own custom blueprint for virtually any document type, including fields with tables of repeating rows and groups of related values. No model training is required.
Every field you extract with a blueprint comes with a confidence score and its location in the source document: the page number and a bounding box around the value. Set a confidence threshold for your workflow, accept values above it automatically, and send the rest for human review. Because each value points back to where it appears, reviewers and auditors can check any result against the original document.
Process mixed document packets, such as loan files, claims files, and onboarding packages, in a single request. Amazon Textract Insights finds where each document begins and ends in a multi-document file, matches each one to the best-fitting blueprint in your project, and extracts data based on fields in that blueprint. Add an optional fallback blueprint for documents that don't match any of the others. To support a new document type, add a blueprint to your project.
Get each document's text, layout elements, tables, and figures in reading order, as natural text, Markdown, or HTML, with tables also available as CSV files. Pass the output straight to a large language model (LLM) or a retrieval-augmented generation (RAG) pipeline without rebuilding the document from JSON blocks. Markdown and HTML keep headings, lists, and tables intact, which helps models understand how the document is organized.
Generate a short summary and a one-sentence description of each document. When a file contains several documents, each one gets its own summary. You can also get a caption for each chart, graph, or image, so the content of visuals is available for search results, generative AI assistants, and downstream applications.
Detect personally identifiable information (PII), including names, addresses, Social Security numbers, and bank account numbers, in your results. You can also write a redacted copy so raw sensitive values don't reach downstream systems. Choose from more than 30 entity types, or detect all of them. Redacted values can be replaced with their entity type, such as [NAME], so the structure of the document stays clear.
Pricing
Open allWith Amazon Textract, you pay only for what you use, with no minimum fees or upfront commitments. For Amazon Textract Core capabilities, you pay for the pages you process. For Amazon Textract Insights capabilities, you pay one price per 1,000 pages for the output type you choose (extraction, extraction with page summaries, or custom extraction), plus optional add-ons.
See Amazon Textract pricing for details, and the FAQ for more about pages and acceptable use.