Runtime Manual/Core Systems/Native Text Embeddings (amber:ai 2.0)

Native Text Embeddings (amber:ai 2.0)

Sub-millisecond text vectorization and high-dimensional cosine similarity directly in V8 without external dependencies

1. Why Native Embeddings in the Runtime?

In Large Language Model applications and Retrieval-Augmented Generation (RAG) pipelines, Text Embedding is the fundamental operation for semantic retrieval, document deduplication, and intent classification. Traditional JavaScript solutions incur heavy operational penalties:

  • Cloud Embedding APIs: 50–200ms round-trip latency, costly for real-time stream filtering or high-throughput batching;
  • Python / PyTorch Child Processes: Hundreds of milliseconds startup latency, high memory overhead, and complex inter-process serialization;
  • Heavyweight Model Files: Multi-hundred-megabyte weights difficult to deploy in lightweight edge agent containers.

Amber v1.3.0 introduces a zero-dependency, native text vectorization engine:

  • Pure Rust Host Engine: Combines subword/n-gram hashing, positional weighting, and GELU activation to generate deterministic dense semantic vectors;
  • Hardware-accelerated L2 Normalization: Outputs unit vectors where the dot product equals exact cosine similarity;
  • Microsecond Latency: Vectorizes sentences in tens of microseconds—over 1,000x faster than network round-trips;
  • Native Integration with amber:vector: Outputs standard Float32Array objects directly consumed by VectorDB.

2. API Reference & Usage

Import embedding utilities from the amber:ai module:

code
import { embed, embedBatch, cosineSimilarity, Tensor } from 'amber:ai';

embed(text, options?)

Computes a dense semantic embedding for an input string:

code
// Defaults to 64-dimensional Float32Array
const vec = embed("High performance JavaScript runtime written in Rust");
console.log(vec instanceof Float32Array); // true
console.log(vec.length); // 64

// Custom dimensions: supports 64, 128, 384
const vec128 = embed("Autonomous agent tool orchestration", { dimensions: 128 });
console.log(vec128.length); // 128

// Return as high-performance Tensor instance
const tensor = embed("Deep learning architectures", { asTensor: true });
console.log(tensor instanceof Tensor); // true

Options (EmbedOptions)

OptionTypeDefaultDescription
dimensionsnumber64Output embedding dimension (64, 128, 384)
normalizebooleantrueApply L2 unit normalization
asTensorbooleanfalseReturn a amber:ai Tensor instance instead of Float32Array

embedBatch(texts, options?)

Generates embeddings for an array of strings in a single call:

code
const texts = [
  "Amber runtime architecture",
  "Zero-copy memory mapping",
  "Homemade Italian pasta recipes"
];

const batch = embedBatch(texts, { dimensions: 64 });
console.log(batch.length); // 3
console.log(batch[0] instanceof Float32Array); // true

cosineSimilarity(a, b)

Computes the cosine similarity between two vector representations in [-1.0, 1.0]:

code
const vA = embed("Concurrency performance and throughput tuning");
const vB = embed("Optimizing server concurrency and system throughput");
const vC = embed("How to bake chocolate brownies in an oven");

const simAB = cosineSimilarity(vA, vB);
const simAC = cosineSimilarity(vA, vC);

console.log("Related similarity:", simAB.toFixed(4));   // ~ 0.75+
console.log("Unrelated similarity:", simAC.toFixed(4)); // ~ 0.15-

3. In-Memory RAG with amber:vector

Combine embed with Amber's built-in VectorDB to build an embedded retrieval pipeline with zero external databases:

code
import { embed } from 'amber:ai';
import { VectorDB } from 'amber:vector';

const db = new VectorDB({ dimensions: 64, metric: 'cosine' });

const docs = [
  { id: '1', text: 'Amber v1.3.0 ships official Model Context Protocol 2.0.' },
  { id: '2', text: 'Virtual Filesystem provides in-memory Copy-on-Write sandbox.' },
  { id: '3', text: 'Embedded SQLite engine supports transactions and prepared queries.' },
];

for (const doc of docs) {
  db.insert(doc.id, embed(doc.text), { text: doc.text });
}

const query = "How does Amber safely run untrusted Agent code?";
const results = db.search(embed(query), { topK: 1 });

console.log("Top Match:", results[0].metadata.text);
PreviousNative AI Engine (amber:ai)NextModern Web Server & Fetch