Crate sevensense_embedding 

sevensense-embedding

Embedding bounded context for 7sense bioacoustics platform.

This crate provides Perch 2.0 ONNX integration for generating 1536-dimensional embeddings from preprocessed audio segments. It handles model loading, inference, normalization, and quantization for efficient storage and retrieval.

Architecture

The crate follows Domain-Driven Design (DDD) principles:

• Domain Layer: Core entities (Embedding, EmbeddingModel) and repository traits
• Application Layer: Services for embedding generation and batch processing
• Infrastructure Layer: ONNX Runtime integration and model management

Usage

use sevensense_embedding::{
    EmbeddingService, ModelManager, ModelConfig,
    domain::Embedding,
};

// Initialize model manager
let config = ModelConfig::default();
let model_manager = ModelManager::new(config)?;

// Create embedding service
let service = EmbeddingService::new(model_manager, 8);

// Generate embedding from spectrogram
let embedding = service.embed_segment(&spectrogram).await?;

Features

• Perch 2.0 Integration: Full support for EfficientNet-B3 bioacoustic embeddings
• Batch Processing: Efficient batch inference with configurable batch sizes
• Model Hot-Swap: Update models without service restart
• Quantization: F16 and INT8 quantization for reduced storage
• Validation: Comprehensive embedding validation (NaN detection, dimension checks)

Re-exports

pub use domain::entities::Embedding;

pub use domain::entities::EmbeddingId;

pub use domain::entities::EmbeddingModel;

pub use domain::entities::EmbeddingMetadata;

pub use domain::entities::StorageTier;

pub use domain::entities::ModelVersion;

pub use domain::entities::InputSpecification;

pub use domain::repository::EmbeddingRepository;

pub use application::services::EmbeddingService;

pub use infrastructure::model_manager::ModelManager;

pub use infrastructure::model_manager::ModelConfig;

pub use infrastructure::onnx_inference::OnnxInference;

Modules

application

Application layer for embedding bounded context.

domain

Domain layer for embedding bounded context.

infrastructure

Infrastructure layer for embedding bounded context.

normalization

Normalization utilities for embedding vectors.

quantization

Quantization utilities for embedding storage optimization.

Enums

EmbeddingError

Unified error type for embedding operations

Constants

EMBEDDING_DIM

Embedding dimension for Perch 2.0 model

MEL_BINS

Mel spectrogram bins for Perch 2.0

MEL_FRAMES

Mel spectrogram frames for Perch 2.0

TARGET_SAMPLE_RATE

Target sample rate for Perch 2.0 (32kHz)

TARGET_WINDOW_SAMPLES

Target window samples (160,000 = 5s at 32kHz)

TARGET_WINDOW_SECONDS

Target window duration in seconds (5s)

VERSION

Crate version information

Type Aliases

Result

Common result type for embedding operations