Skip to content

Assisted Business Document Type Suggestions

SageMaker models SageMaker model

Give a reviewer two possible types for already extracted English, Spanish or Portuguese business text. You compare both suggestions with the source and explicitly choose a workflow or reject them.

Use it after text extraction when you organize invoice, purchasing or document-intake work. It suggests purchase order, delivery note, bank statement, utility bill, insurance policy, invoice or other business correspondence. The caller establishes that the text is one supported document. The other label means business correspondence; it does not validate arbitrary unknown inputs.

Delivery
SageMaker model package in your own AWS account

What you receive

Version 1.0.0 returns two ranked labels, stable caller IDs, analyzed token/code-point counts and mandatory review/selection flags. Rank is not confidence. The offline CPU runtime includes the frozen multilingual E5 encoder and an original ridge head. It needs no runtime model download.

It accepts text only. It does not perform OCR, extract fields, verify authenticity, detect unknown or combined documents, or automatically route a document. Out-of-scope text still receives suggestions. A reviewer must reject unsuitable input.

Software and AWS costs

The software is USD 0.50 per instance-hour on ml.c5.xlarge for real-time or Batch Transform, with a 14-day software trial. AWS compute, storage, transfer and other services are separate, including during the trial. The current Marketplace offer is authoritative. Delete endpoints when finished; closing a notebook does not stop billing.

Start here

Marketplace listing · Public notebook and examples.

Watch the workflow

English desktop walkthrough · Vertical Short.

The videos show actual model calls on fictional text in an illustrative UI. The demonstration choices do not alter evaluation labels or trigger downstream actions. AI narration is disclosed.

Upstream project

The frozen encoder is intfloat/multilingual-e5-small, under MIT. Software Sushi's original serving code and head use Apache-2.0. The public examples repository contains the SDK notebook and fictional inputs/recorded outputs; it does not distribute encoder weights.

Last updated November 16, 2026 · Suggest an improvement