Source-Grounded Business Summaries
Software Sushi

Available on AWS Marketplace · Version 0.1.0

Buyer usage guide

The notebook accepts the regional package ARN from your own subscription and verifies the exact release-approved image digest. Set GROUNDED_SUMMARY_MODEL_PACKAGE_ARN to the ARN supplied after Continue to configuration in AWS Marketplace. Default Run All makes no AWS calls. After reviewing costs, permissions and cleanup, set RUN_DEMO = True to enable real-time deployment and RUN_BATCH = True for the optional batch example. Keep RUN_DEMO enabled during cleanup. A buyer-account subscription transaction has not been rehearsed by the seller. Check the release status and deploy only after an active AWS Marketplace subscription. Run in us-east-1 using your own SageMaker notebook role with boto3 installed. Outside a role session, configure SAGEMAKER_EXECUTION_ROLE_ARN. The notebook creates a fresh private encrypted evidence bucket or verifies the private bucket supplied through SAGEMAKER_DEMO_BUCKET. Existing buckets are checked without reconfiguration.

Request shape

Send one JSON object per invocation with Content-Type: application/json:

{
  "language": "en",
  "segments": [
    {"id": "s1", "text": "Maya Chen will send the signed transition plan by 2027-06-18."}
  ]
}

For Spanish or Portuguese, set language to es or pt and provide text in that language. For example, Spanish text may say Lucía Mena enviará el cronograma firmado antes del 2027-06-18. Portuguese text may say João Luís enviará o cronograma assinado até 2027-06-18. Dates qualify as deadlines only when the source supplies a valid ISO calendar date (YYYY-MM-DD). Relative timing such as “next week” is not converted into a date.

Each segment requires a unique, nonempty id and nonempty text. The request supports 1–64 segments, up to 2,000 characters per segment, and 16,000 combined segment-ID and text characters. The HTTP body must not exceed 96 KiB. The runtime checks that the actual input fits 3,072 tokens and caps generated output at 1,024 tokens. Oversized input is rejected; there is no silent truncation. Only one structured record is processed per invocation.

JSON Lines and batch

Use Content-Type: application/jsonlines with one complete request object per line. Configure line-split batch processing with one record per invocation. The notebook and public batch example contain three fictional EN/ES/PT records. Each is sent in a separate invocation. Use SingleRecord, SplitType=Line and AssembleWith=Line; real-time and batch use the same inference route. The notebook includes deployment, invocation and cleanup cells.

Response fields

The response includes:

Evidence quote is the full selected source segment. Offsets use Unicode code points: start is zero and end is the segment's code-point length. The model selects segment IDs; the backend copies segment text. A literal match does not establish semantic entailment. Confirm every summary and action against the original document.

Scope and deployment

The model is for human-reviewed drafts. It does not use tools, create tasks, send messages, or make decisions. It accepts text only and does not perform OCR, ASR, or image processing. Validate results on your own documents and language mix; development fixtures are fictional and have no human/native acceptance.

The qualified types are ml.g6e.4xlarge for real-time and ml.g6e.2xlarge for batch, one 48 GB GPU each. Your account needs the corresponding quotas; quota approval does not guarantee AWS capacity. The runtime passed a 60-record development evaluation on 4xlarge and actual EN/ES/PT certification batch on 2xlarge. These checks are not independent customer accuracy. A measured cold deployment took about 21 minutes, which is outside the per-request latency measurement. Allow up to 40 minutes in the notebook waiter and run cleanup after failures.

Support: info@softwaresushi.com. Include the product version, AWS region, instance type, and request ID. Do not send source documents, private data, or credentials. No response-time SLA is advertised. Consult the active offer for software pricing and trial terms. AWS infrastructure charges are separate.

Hosting costs

The AWS Price List API on 13 November 2026 quotes USD 3.7553 per hour for on-demand ml.g6e.4xlarge Hosting in US East (N. Virginia), excluding storage, transfer and software charges. This is a hosting SKU, not a Studio price. At the USD 0.50/hour software rate, compute plus software would be USD 4.2553/hour before those other charges. During a software trial, AWS infrastructure is still billed. A separate Batch Transform SKU quotes USD 2.8026/hour for ml.g6e.2xlarge in the same region, or USD 3.3026/hour including the software fee before storage and transfer. Regional availability must be checked separately. Stop/delete temporary endpoints promptly after testing. Consult the active Marketplace offer and release status before deploying.