@vncsleal/prisml-runtime

ONNX Runtime prediction engine for PrisML - compiler-first machine learning for TypeScript + Prisma.

Features

• In-process ONNX inference
• Schema drift detection
• Single and batch predictions
• Type-safe error handling
• Synchronous, deterministic execution

Installation

npm install @vncsleal/prisml-runtime

Usage

import { PredictionSession, hashPrismaSchema } from '@vncsleal/prisml-runtime';

const session = new PredictionSession();
const schemaHash = hashPrismaSchema(schemaContent);

// Initialize model
await session.initializeModel(
  './.prisml/model.metadata.json',
  './.prisml/model.onnx',
  schemaHash
);

// Single prediction
const result = await session.predict('modelName', entity, {
  price: (e) => e.price,
  stock: (e) => e.stock,
});

console.log(result.prediction); // Predicted value

// Batch predictions
const results = await session.predictBatch('modelName', entities, resolvers);

Error Handling

All errors extend PrisMLError with structured context:

import { SchemaDriftError, UnseenCategoryError } from '@vncsleal/prisml-runtime';

try {
  await session.predict('model', entity, resolvers);
} catch (error) {
  if (error instanceof SchemaDriftError) {
    // Handle schema mismatch
  }
}

Documentation

See main documentation and user guide.

License

MIT © Vinicius Leal