Cohere Embedding Provider

The Cohere embedding provider implements the embedding.Embedder interface using Cohere’s v2 Embed API. It supports input type differentiation between documents and queries, which is critical for asymmetric search scenarios.

Choose Cohere embeddings when building search systems that benefit from asymmetric retrieval (different embeddings for documents vs. queries). Cohere also offers multilingual models that embed 100+ languages into a shared vector space, making them ideal for cross-language search applications.

Installation

go get github.com/lookatitude/beluga-ai/rag/embedding/providers/cohere

Quick Start

package main

import (
    "context"
    "fmt"
    "log"
    "os"

    "github.com/lookatitude/beluga-ai/config"
    "github.com/lookatitude/beluga-ai/rag/embedding"
    _ "github.com/lookatitude/beluga-ai/rag/embedding/providers/cohere"
)

func main() {
    emb, err := embedding.New("cohere", config.ProviderConfig{
        APIKey: os.Getenv("COHERE_API_KEY"),
    })
    if err != nil {
        log.Fatal(err)
    }

    ctx := context.Background()
    vectors, err := emb.Embed(ctx, []string{
        "Beluga AI provides a unified embedding interface",
        "Vector databases store high-dimensional embeddings",
    })
    if err != nil {
        log.Fatal(err)
    }

    fmt.Printf("Embedded %d texts, dimensions: %d\n", len(vectors), emb.Dimensions())
}

Configuration

Parameter	Type	Default	Description
APIKey	string	(required)	Cohere API key
Model	string	embed-english-v3.0	Embedding model name
BaseURL	string	https://api.cohere.com/v2	API base URL
Timeout	time.Duration	0 (no timeout)	Request timeout
Options["input_type"]	string	search_document	Input type for embedding
Options["dimensions"]	float64	Model-dependent	Override vector dimensions

Supported Models

Model	Default Dimensions
embed-english-v3.0	1024
embed-multilingual-v3.0	1024
embed-english-light-v3.0	384
embed-multilingual-light-v3.0	384
embed-english-v2.0	4096

Input Types

Cohere embedding models differentiate between document and query embeddings. Set the input_type option to control this behavior:

• search_document (default) — Use when embedding documents for storage
• search_query — Use when embedding queries for retrieval
• classification — Use for classification tasks
• clustering — Use for clustering tasks

// Embed documents for storage
docEmb, err := embedding.New("cohere", config.ProviderConfig{
    APIKey: os.Getenv("COHERE_API_KEY"),
    Options: map[string]any{
        "input_type": "search_document",
    },
})

// Embed queries for retrieval
queryEmb, err := embedding.New("cohere", config.ProviderConfig{
    APIKey: os.Getenv("COHERE_API_KEY"),
    Options: map[string]any{
        "input_type": "search_query",
    },
})

Direct Construction

import (
    cohereemb "github.com/lookatitude/beluga-ai/rag/embedding/providers/cohere"
)

emb, err := cohereemb.New(config.ProviderConfig{
    APIKey: os.Getenv("COHERE_API_KEY"),
    Model:  "embed-multilingual-v3.0",
    Options: map[string]any{
        "input_type": "search_document",
    },
})
if err != nil {
    log.Fatal(err)
}

Multilingual Support

Cohere’s multilingual models support 100+ languages in a shared embedding space:

emb, err := embedding.New("cohere", config.ProviderConfig{
    APIKey: os.Getenv("COHERE_API_KEY"),
    Model:  "embed-multilingual-v3.0",
})
if err != nil {
    log.Fatal(err)
}

vectors, err := emb.Embed(ctx, []string{
    "Hello world",           // English
    "Bonjour le monde",      // French
    "Hola mundo",            // Spanish
})