Software SushiAssisted Business Document Type Suggestions

Available on AWS Marketplace · Version 1.0.0

Reviewing document type suggestions

Deploy through your AWS Marketplace subscription and use the regional model-package ARN shown in the product configuration. The seller's source qualification ARN is not a buyer subscription ARN. The complete downloadable notebook defaults to no AWS calls; explicitly enable deployment only after reading its cost, access and cleanup instructions.

  1. Extract the text with your own authorized PDF/OCR component, preserving its actual content. This service receives no file and does not perform extraction.
  2. Establish that the text comes from a single supported business document. Combined, unrelated or insufficient text can still receive two suggestions.
  3. Send one JSON envelope to the /invocations endpoint. The complete text must fit the token limit; do not silently clip a document to obtain a label.
  4. Present both suggestions to a reviewer. Require an explicit choice or rejection before invoking a downstream processor. Rank is not confidence.
  5. Evaluate your own authorized examples before relying on the suggestions. Real customer-document accuracy is unmeasured; existing diagnostics use model-authored examples and provisional labels.

Example request structure:

{"instances":[{"id":"business-review-1","text":"Your already extracted source text","filename":"optional-metadata.txt"}]}

The response includes predictions, preserving each ID. Each prediction has two distinct suggestions with rank and document_type, mandatory needs_review and selection_required, input_tokens, analyzed_codepoints, whole_document_coverage=unknown, truncated=false, unknown_mixed_detection=not_provided and validation_basis=generated_only. The API declares zero independent acceptance rows. Publication status is recorded in the product listing rather than asserted by inference results. It returns no definitive selected document_type or probability. The exact image passed private real-time and full-cohort Batch Transform checks on ml.c5.xlarge in us-east-1. See the model card for actual workload/latency limits; these are not customer accuracy or service-level guarantees.

HTTP limits and handling:

Status Action
400 Correct the complete malformed/invalid envelope; no partial predictions
413 Reduce request size or batch count while preserving complete documents
415 Use JSON or one-envelope-per-line JSON Lines without content encoding
422 Complete text exceeds 256 encoder tokens; use a separately designed workflow
429 Back off and retry the same request; inference is stateless
500 Retry boundedly or surface the service failure for review

JSON Lines requires exactly one complete envelope per invocation. Qualified Batch Transform configuration uses one ml.c5.xlarge, line splitting, SingleRecord, line assembly, MaxConcurrentTransforms=1 and MaxPayloadInMB=1. Each envelope still obeys the smaller 256 KiB/32-record runtime bounds. MultiRecord is unsupported. Set a task-owned S3 input/output prefix and execution role. The notebook documents real-time setup, invocation and exact-resource cleanup; delete the endpoint when finished. Stopping a notebook does not stop endpoint billing. Buyer-controlled account access, encryption, retention, authorized source use and downstream decisions remain part of your application design.

Complete SDK examples

The public GitHub notebook includes both real-time and optional Batch Transform examples with fictional ES/PT/EN inputs, actual certified-image output samples, explicit opt-in, cost notes and cleanup. Batch uses your existing authorized us-east-1 S3 bucket; it creates no bucket or role and retains its unique example prefix for inspection. Default Run All makes no AWS calls. Use a current boto3 release supporting conditional S3 PutObject (IfNoneMatch) before opting in.

Public examples and notebook.