Amazon Textract Insights Now Generally Available: Extract Document Fields in Natural Language, With Confidence Scores and Page Grounding, Up to 3,000 Pages per File
AWS announced general availability of Amazon Textract Insights, which lets you define fields in natural language and process mixed document packets, with a confidence score and page grounding for each value. It supports six languages, and is not yet available in Tokyo or Zurich.
News date: About 6 min read
Contents
AWS says Amazon Textract Insights has reached general availability. It is a generative AI capability within Amazon Textract, delivered through a new API. Users describe in natural language which fields to extract and get structured results with confidence scores and page grounding. AWS cites mixed document packets such as mortgage, claims and onboarding packets as the typical scenario (all per AWS).
How does Amazon Textract Insights differ from the traditional approach?
AWS notes that teams usually take one of two approaches. The first is to assemble a custom pipeline of OCR (optical character recognition, which turns text in scanned images into machine-readable text), document splitting, classification, foundation model prompting and review steps. The second is to use a general-purpose large language model, which can give fluent answers but may not reliably indicate the source page or provide a confidence score. AWS says Textract Insights consolidates the whole flow into a single asynchronous request on Textract, meaning you submit it, it processes in the background, and you retrieve the result when it finishes.
According to AWS, users define what to extract with a "blueprint" (think of it as an extraction template). It is a JSON Schema, the standard format for describing the structure of data fields. Each field has a natural-language description and an "inference type" that determines whether the value is read verbatim from the page or inferred from the document's content. AWS's public blueprints cover W-2, Form 1040, Form 1098, passports and birth certificates, among others, and users can write their own.
How are mixed document packets handled, and which formats and languages are supported?
To process a mixed packet, users first add the blueprints for each document type to the same project. They then submit a single file (up to 3,000 pages) from their own Amazon S3 bucket (AWS's cloud file storage service) with splitting enabled. AWS says the service splits the file into logical sub-documents, matches each to a suitable blueprint, extracts the fields, and returns a confidence score and grounding for each value.
AWS gives an example: a 26-page mortgage application PDF containing a W-2, pay stubs, bank statements and handwritten loan officer notes. A single StartDocumentInsights call is enough, and the results are written back to the user's S3. Each value comes with a confidence score, plus its source page and bounding box (the box coordinates marking where the value sits on the page).
| Item | Textract Insights | Textract Core update |
|---|---|---|
| Nature | New generative AI capability, delivered through a new API | Update to the existing reading engine; existing APIs unchanged |
| Languages | English, French, German, Italian, Portuguese, Spanish | DetectDocumentText and AnalyzeDocument expand to 40 languages |
| Regions | Not yet available in Tokyo or Zurich | Adds Tokyo and Zurich, for 18 regions in total |
| Processing location | Inference automatically routed to a region within the user's geography | Processed in the region where the call is made |
| Input and pricing | PDF, PNG, JPEG, TIFF; per-page pricing, no minimum commitment | Not stated in the announcement |
How do confidence scores and grounding support auditing and automation?
AWS explains that Insights returns two kinds of fields: explicit fields read directly from the page, and inferred fields derived from the content. Both come with a confidence score and the page and coordinates the value was taken from. Reviewers or auditors can therefore check any result against the original document.
AWS quotes an assessment from Vijay Agrawal, Vice President of Information Systems at Symbeo (a CorVel company): teams can describe the fields they need in plain language, review only the lower-confidence fields based on confidence scores, and he looks forward to exploring the document splitting feature. This is a customer statement published by the vendor and has not been independently verified.
Data security and how to get started
AWS says projects and blueprints are encrypted at rest by default with AWS-owned keys, meaning data is encrypted while stored. To use your own AWS KMS (AWS's key management service) customer managed key, specify it in EncryptionConfiguration when calling CreateProject or CreateBlueprint. Intermediate processing data for analysis jobs is also encrypted, is temporary, and is deleted when the job completes. A project's key does not apply to its jobs.
- Define your needs: start from AWS's public blueprints, or describe your own fields in natural language.
- Create a project: add the blueprints to the project.
- Submit documents: send a single document or a mixed packet from Amazon S3 with one API call.
- Review results: accept values that meet your confidence threshold, send the rest to human review, and tune the threshold with your own representative sample.
Frequently asked questions
Is Amazon Textract Insights currently a preview or a production release?
AWS states in its announcement that it has reached general availability, meaning a production release. Region and language limits still apply, however; for example, Tokyo and Zurich are not yet supported.
Do I need to train my own model?
No. AWS says users simply describe fields in natural language, with no model training required.
Will the existing Textract APIs be affected?
AWS says the Textract Core updates require no changes from existing customers, and existing APIs remain unchanged.
How is it priced?
AWS says pricing is per page rather than per token, with no minimum commitment. The announcement does not list specific rates, so see the Amazon Textract pricing page.
Can it be used in Taiwan?
The AWS announcement does not provide a full list of availability by region or for Taiwan. It only states that Insights is not yet available in Tokyo and Zurich, so rely on the AWS documentation.
Found this guide helpful? Share it with someone who needs it
Claude Code 2.1.330: Enterprise Admin Console Sign-In, easel Chart Module, and Scheduled Tasks Extended to 30 Days
Anthropic's Claude Code 2.1.330 changelog lists Enterprise Admin Console sign-in, a built-in easel chart module, scheduled tasks of up to 30 days, and stricter approval for rm commands.
AWS Quick adds chat-built flows, multi-agent answers and Meta ads/Amplitude connectors, available today in all Quick regions
AWS Distributed KV Caching Agent Skill for Self-Hosted LLMs: AWS Says Cost Drops Up to 40% and Time to First Token Up to 80%
Amazon SageMaker HyperPod Slurm Adds Cluster Extensions: Managed Observability and User Management, Available in All HyperPod Regions
Sources
Compiled from the references and verification dates listed above; fares, schedules, product features, and terms can change, so please confirm the latest details on the relevant official website before travelling, booking, or purchasing.