web-decision

[email protected] supports three browser-local decision models:

• kev-0.6b-q4f16 uses TVM WebGPU and requires shader-f16.
• experimental-tamev-nano-tinybert-int8 uses ONNX Runtime WebAssembly on CPU with one inference thread.
• experimental-tamev-micro-minilm-int8 uses ONNX Runtime WebAssembly on CPU with one inference thread.

Each model uses its own pinned asset directory. The package contains JavaScript, declarations, runtime dependencies and notices; it does not contain model weights.

Install

npm install [email protected]

Kev 0.6B

Host the exact Kev asset pack at a URL your application can fetch, then pass its directory URL to the loader:

import {loadDecisionModel} from 'web-decision';

const model = await loadDecisionModel({
  assetsUrl: 'https://your-origin.example/models/kev-0.6b-q4f16/'
});

try {
  const result = await model.decide({
    state: 'A customer requests a refund for a damaged item.',
    questions: [{instructions: 'Should the refund be approved?', options: ['Approve', 'Decline']}]
  });
  console.log(result.answers[0]?.selected);
} finally {
  await model.dispose();
}

The Kev loader verifies its pinned manifest and asset hashes. The exact 0.1.0 archive passed real WebGPU tests in Chromium, Firefox and WebKit on one Apple M4 Pro. Those results apply to that archive and test device.

TAMEV Nano TinyBERT

Copy the four files from the pinned TAMEV Nano model revision to a directory your application can fetch. The required files total 15,585,045 bytes: model_int8.onnx, tokenizer.json, tokenizer_config.json and config.json.

import {loadDecisionModel} from 'web-decision';

const model = await loadDecisionModel({
  model: 'experimental-tamev-nano-tinybert-int8',
  assetsUrl: 'https://your-origin.example/models/tamev-nano-tinybert/'
});

try {
  const result = await model.systemOne(
    {payment: 'duplicate card charge'},
    {
      route: {
        type: 'choice',
        instructions: 'Select the handling queue',
        criteria: {billing: 'review a duplicate charge', fraud: 'investigate unauthorized use'}
      },
      urgent: {type: 'noul', instructions: 'Does the customer need immediate help?'},
      severity: {type: 'score', instructions: 'Rate the impact', criteria: ['low', 'medium', 'high']}
    }
  );
  console.log(result.answers.route?.choice);
} finally {
  await model.dispose();
}

The TAMEV adapter verifies each downloaded file against its pinned byte length and SHA-256 hash. It preserves option and question order, supports Choice, Noul and Score, and uses the model's calibration for option counts with a configured calibration entry. Noul returns the model's yes probability; it does not abstain.

TAMEV Nano runs through ONNX Runtime WebAssembly, not WebGPU. Its context rows allow 128 tokenizer tokens and each option row allows 48; excess text is right-truncated and reported. The package imposes no fixed option-count or question-count ceiling. Questions run sequentially and practical request size depends on available device memory.

The upstream model card reports its own quality metrics and warns that external tasks may differ. web-decision has verified browser execution and input/output behavior; it does not independently validate decision quality.

TAMEV Micro MiniLM

Copy the four files from the pinned TAMEV Micro model revision to a directory your application can fetch. The required files total 24,161,003 bytes: model_int8.onnx, tokenizer.json, tokenizer_config.json and config.json.

import {loadDecisionModel} from 'web-decision';

const model = await loadDecisionModel({
  model: 'experimental-tamev-micro-minilm-int8',
  assetsUrl: 'https://your-origin.example/models/tamev-micro-minilm/'
});

try {
  const result = await model.systemOne(
    {payment: 'duplicate card charge'},
    {
      route: {
        type: 'choice',
        instructions: 'Select the handling queue',
        criteria: {billing: 'review a duplicate charge', fraud: 'investigate unauthorized use'}
      },
      urgent: {type: 'noul', instructions: 'Does the customer need immediate help?'},
      severity: {type: 'score', instructions: 'Rate the impact', criteria: ['low', 'medium', 'high']}
    }
  );
  console.log(result.answers.route?.choice);
} finally {
  await model.dispose();
}

The adapter verifies each downloaded file against its pinned byte length and SHA-256 hash. It preserves option and question order, supports Choice, Noul and Score, and applies the model's exported probabilities plus its configured per-option-count calibration to decide(). Noul returns the model's yes probability; it does not abstain.

TAMEV Micro runs through ONNX Runtime WebAssembly, not WebGPU. Its context rows allow 128 tokenizer tokens and each option row allows 48; excess text is right-truncated and reported. The package imposes no fixed option-count or question-count ceiling. Questions run sequentially and practical request size depends on available device memory. The upstream model card's quality figures are not independently reproduced by this project.

Asset hosting and safety

The application must host model files under the supplied assetsUrl and allow browser requests from its origin. Keep the URL in trusted application configuration: the loader fetches assets from it. Decision inputs are not sent to the asset host. Model assets are not part of the npm archive; the package includes the TAMEV model notices and license files.

The Kev route also excludes its 14 inference assets. Its pinned manifest totals 349,103,828 bytes. KEV06-NOTICE and KEV-NOTICE cover the route and adapted code; review the terms that apply to the separately hosted asset files.

API

• loadDecisionModel({assetsUrl}) loads Kev 0.6B by default.
• loadDecisionModel({model: 'experimental-tamev-nano-tinybert-int8', assetsUrl}) loads TAMEV Nano TinyBERT.
• loadDecisionModel({model: 'experimental-tamev-micro-minilm-int8', assetsUrl}) loads TAMEV Micro MiniLM.
• All three sessions provide decide(), releaseCachedMemory() and dispose(). They expose typed systemOne() results for their supported protocols.
• Use listDecisionModels(), isDecisionModel() and describeModel() to inspect the three pinned routes.
• Use clearModelCache() to clear this library's asset cache for the current origin.

The package code is licensed under Apache-2.0. Model terms and runtime asset terms remain separate; see the included notices.