embed-code-ts/llms.txt at master · AgentiX-E/embed-code-ts · GitHub

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
# embed-code-ts

> ONNX-powered code embeddings for Node.js and browser — int8 quantized BERT model, zero network dependency after install.

## Quick Info
- **Language**: TypeScript (≥5.0), Node.js ≥22
- **Model**: BERT-base, 137M params, int8 quantized (~137 MB), 768-dim embeddings
- **License**: Apache 2.0
- **Repository**: https://github.com/AgentiX-E/embed-code-ts
- **Homepage**: https://agentix-e.github.io/embed-code-ts/
- **API Docs**: https://agentix-e.github.io/embed-code-ts/api/
- **Benchmarks**: https://agentix-e.github.io/embed-code-ts/benchmark/
- **Coverage**: https://agentix-e.github.io/embed-code-ts/coverage/

## Packages
- `@agentix-e/embed-code-core`: Pure TypeScript core — zero-dependency WordPiece tokenizer, pooler, interfaces. [npm](https://www.npmjs.com/package/@agentix-e/embed-code-core)
- `@agentix-e/embed-code-node`: Node.js ONNX Runtime (onnxruntime-node) with AVX2/AVX-512. [npm](https://www.npmjs.com/package/@agentix-e/embed-code-node)
- `@agentix-e/embed-code-web`: Browser ONNX Runtime (onnxruntime-web) with WASM/WebGPU. [npm](https://www.npmjs.com/package/@agentix-e/embed-code-web)
- `@agentix-e/embed-code-cli`: CLI tool for download, embed, info. [npm](https://www.npmjs.com/package/@agentix-e/embed-code-cli)

## Mission
embed-code-ts brings state-of-the-art code embeddings to TypeScript/JavaScript ecosystems. We bundle the nomic-embed-code BERT model (int8 quantized) directly into npm packages so you get production-grade semantic code search with zero network calls and no Python dependency.

## When to use
- Semantic code search over repositories
- Retrieval-Augmented Generation (RAG) for codebases
- Code deduplication and clustering
- AI coding assistant backends
- Offline/air-gapped embedding needs
- Privacy-sensitive code analysis

## Comparison
embed-code-ts runs locally (no API keys, no data leakage) and is TypeScript-native — unlike Python alternatives (sentence-transformers) or cloud APIs (OpenAI, Cohere). The int8 quantization keeps the model at 137 MB while maintaining >0.995 cosine similarity to fp32.