onnx/README.md

ONNX exports

Source: Shuu12121/NightJar-CodeSearch-Embedding

File	Precision
model.onnx	FP32
model_quantized.onnx	Dynamic INT8 (constant MatMul and Gather weights)

Inputs: input_ids and attention_mask, int64 [batch, sequence]. Output: last_hidden_state. Take the first token, then L2-normalize to obtain 768-dimensional embeddings. Maximum input: 1024 tokens. Use the tokenizer/configuration at the repository root. Query/document prefixes are not required.

Node.js short-input example (Transformers.js 4.2.0):

import { pipeline } from '@huggingface/transformers';
const extractor = await pipeline('feature-extraction', 'Shuu12121/NightJar-CodeSearch-Embedding', {
  dtype: 'q8', device: 'cpu',
});
const output = await extractor('parse a JSON string', { pooling: 'cls', normalize: true });
console.log(output.dims); // [1, 768]
await extractor.dispose();

For long inputs, truncate the content to 1022 tokens before adding CLS and SEP. Transformers.js 4.2.0's default truncation slices the final sequence and can remove SEP; preserve both special tokens for parity with the Python tokenizer.

manifest.json records the source commit, export settings, validation results and SHA256 checksums. These exports use opset 17. The original model's license and usage terms apply. INT8 output embeddings remain FP32.