lnmp-embedding

Native vector embedding support for the LNMP protocol.

FID Registry: Embedding FIDs are defined in the Standard range (512-767) in registry/fids.yaml.

This crate provides the core data structures and logic for handling vector embeddings within LNMP, enabling efficient storage, transport, and processing of high-dimensional vectors for AI and ML applications.

Features

• Multiple Data Types: Supports F32, F16, I8, U8, and Binary embeddings.
• Similarity Metrics: Built-in calculation for Cosine Similarity, Euclidean Distance, and Dot Product.
• ** efficient Serialization**: Optimized binary format for minimal overhead.

Usage

use lnmp_embedding::{Vector, EmbeddingType, SimilarityMetric};

// Create a vector from f32 data
let v1 = Vector::from_f32(vec![1.0, 0.0, 0.0]);
let v2 = Vector::from_f32(vec![0.0, 1.0, 0.0]);

// Calculate similarity
let similarity = v1.similarity(&v2, SimilarityMetric::Cosine).unwrap();
assert_eq!(similarity, 0.0);

Examples

This crate includes several examples in the examples/ directory:

• delta_compute: Computing embedding deltas
• batch_ops: Batch vector operations (similarity, etc.)

Run examples with:

cargo run --example delta_compute -p lnmp-embedding
cargo run --example batch_ops -p lnmp-embedding

Integration

This crate is designed to be used with lnmp-core and the LNMP SDKs. It is the underlying implementation for the Embedding mode (0x06) and LnmpValue::Embedding variant.