Software SushiAssisted Business Document Type Suggestions

Available on AWS Marketplace · Version 1.0.0

Assisted Business Document Type Suggestions (ES/PT/EN)

Version 1.0.0 · Available on AWS Marketplace.

Suggest two possible types for a person reviewing already extracted English, Spanish or Portuguese business text. Useful after text extraction when a reviewer needs to choose a downstream workflow, such as invoice extraction, order intake or receipt handling. Selection and subsequent actions remain the caller's job.

Intended input and output

Input must be one document already known to belong to seven supported categories: purchase order, delivery note, bank statement, utility bill, insurance policy, invoice, or other business correspondence. The endpoint returns two ranked labels, caller ID, analyzed code points/tokens and mandatory selection/review flags. Rank does not express a calibrated probability or a final routing decision.

The service accepts text only, not PDF/image/audio or whole-file parsing. Whole-document coverage is unknown. It does not validate whether an input belongs to the supported scope, detect combined/unknown documents, verify authenticity, extract fields, assess insurance eligibility, or decide payments or credit. Out-of-scope text can receive plausible-looking suggestions. A reviewer must reject unsuitable inputs rather than selecting one of the two suggestions.

Model and data provenance

Component Frozen provenance
Encoder/tokenizer intfloat/multilingual-e5-small, revision 614241f622f53c4eeff9890bdc4f31cfecc418b3, MIT
Encoder SHA256 ca456c06b3a9505ddfd9131408916dd79290368331e7d76bb621f1cba6bc8665
Tokenizer SHA256 0b44a9d7b51c3c62626640cda0e2c2f70fdacdc25bbbd68038369d14ebdf4c39
Ridge head SHA256 1e438bdae0f1fd978e21cd6e58aaea7752c56771ee158edd6e4a4fab569ed87f
Fitting 3,600 original fictional TRAIN rows plus 324 new Luna-authored TRAIN rows; source-balanced family means, fixed ridge penalty .01
Ranking Project the unchanged nine-column head onto its seven supported columns; sort descending raw values with alphabetical ties; return two labels without scores
New final dataset SHA256 050149c2134bedf108767667c3793f3b28342b777615a7de9c3c15b469f5b90f
Software and original head Apache-2.0; dependency notices retained in the image

IDs, filenames, language metadata, gold labels and reviewer judgments are not features. No final/challenge examples were fitted. The head was selected using exposed generated DEV results; the fresh generated final protocol was committed before its authors began. The source ledger and pinned upstream card are in the parent research directory; this card does not assert knowledge of every upstream pretraining example.

All project document examples are fictional/model-authored. Separate authors share the model, taxonomy and brief. Exact normalized duplicate checks passed against all prior project corpora; they do not establish semantic independence. There are zero native documents and zero independent human acceptance labels. Two blinded local AI reviewers provide label diagnostics, not human adjudication.

Across all 324 final excerpts, Qwen agrees with the author on 322 and Gemma on 222. The two reviewers agree on 224 rows, including two supported examples where both prefer invoice over the author's utility_bill; the two suggestions include both alternatives in each case. One hundred reviewer disagreements remain unresolved. No author label was changed and no row was excluded.

Generated final diagnostic

The fixed 0.6 prototype was evaluated once on 324 newly authored excerpts through the exact offline HTTP image: 84 supported examples and 24 out-of-scope controls per language, 12 examples per type/language. Every row is retained, all inputs were admitted, and none were relabeled or excluded after inference.

Language First suggestion matches provisional author Author label among two suggestions
English 81/84 (96.4%) 84/84
Spanish 79/84 (94.0%) 84/84
Portuguese 80/84 (95.2%) 84/84

The 1.0 release API preserves its head, ranking, input bounds and label behavior; it removes prototype publication-state flags and keeps generated-only/native-zero and mandatory review/selection metadata. Complete-cohort serving parity is a separate engineering gate, not a repeated final-quality claim.

The declared generated diagnostic floors pass in every language/type. These counts describe this dependent generated corpus only. They are not estimates of customer-document accuracy, and no representative statistical confidence interval is claimed. All 72 unknown/mixed controls receive suggestions and remain outside the product claim; this does not count as successful detection. All 12 supported first-suggestion disagreements and AI label-review disagreements remain visible in the evidence. Model performance on real customer documents is unmeasured.

Sensitivity to reviewer labels matters: on the same fixed 84 author-supported rows per language, the two suggestions contain Qwen's label on 84/84 in each language, but Gemma's on EN 81/84, ES 76/84 and PT 71/84. These results are retained alongside the author-label table; the weaker reviewer result must not be hidden behind the 100% author-label coverage. There is no adjudicated gold set proving which model is correct. The declared diagnostic floors are against provisional author labels only, as frozen before evaluation.

Runtime, privacy and qualification

Bounds: 256 KiB request, 1–32 records, 16,000 code points/text, at most 256 complete encoder tokens including prefix/special tokens. Overflow rejects the entire batch, without truncation. Metadata is never fetched or used as a path/URL. One inference runs at a time; contention returns 429. Errors are sanitized; request text and paths are not logged. All weights are baked into the image.

Actual 0.6 SER9 execution used no network, UID 1000, read-only root, no-new-privileges, two CPU and 3 GiB. HTTP/JSON Lines, filename invariance, whole-batch rejection, recovery and SIGTERM passed. That image reports zero Trivy findings; the exact archive/config chain and raw scan are retained. This is a point-in-time scan, not a guarantee against all vulnerabilities or AWS certification. The 1.0 image separately reproduced all 567 exposed reference suggestions with zero differences, including all 126 controls. Its four executed runtime source hashes match frozen Git commit 48ff1c31f4c3310e316d01efcf8afb294675e68b. Trivy and the current ECR scan both report zero findings. SageMaker rejected the initial OCI manifest before creating compute; conversion to Docker schema 2 preserved every config/layer byte. Qualified registry digest: sha256:58d7c9ad8b9f56f1ccb2f313cc9380327a7e45d91a5af4dd972b7ea34d9da89b.

Actual private AWS checks in us-east-1, one ml.c5.xlarge:

AWS Marketplace made this assisted product public on 16 November 2026. Software is USD 0.50 per ml.c5.xlarge instance-hour, with a 14-day software trial; AWS infrastructure is separate. The released offer is authoritative. The actual SDK notebook passed real-time and three-language batch inference in the seller account through the product-associated package, then verified cleanup. This is not an independent buyer checkout or entitlement test. All technical qualification was in us-east-1; other listed regions have not been benchmarked. Actual AWS checks do not establish native customer-document accuracy.