Cohere Embeddings

The Cohere embeddings provider converts text into vectors using Cohere’s embed-v3.0 models. These models offer strong multilingual support and efficient embedding dimensions.

Installation

npm install @youcraft/recall-embeddings-cohere

Usage

import { createMemory } from '@youcraft/recall'
import { cohereEmbeddings } from '@youcraft/recall-embeddings-cohere'

const embeddings = cohereEmbeddings({
  apiKey: process.env.COHERE_API_KEY!,
})

const memory = createMemory({ db, embeddings, extractor })

Configuration

Option	Type	Default	Description
apiKey	string	required	Your Cohere API key
model	string	"embed-english-light-v3.0"	Embedding model to use
inputType	string	"search_document"	Input type for embeddings

Models

Model	Dimensions	Description
embed-english-v3.0	1024	High quality English embeddings
embed-english-light-v3.0	384	Fast, efficient English embeddings
embed-multilingual-v3.0	1024	High quality multilingual (100+ languages)
embed-multilingual-light-v3.0	384	Fast multilingual embeddings

Examples

High quality English

const embeddings = cohereEmbeddings({
  apiKey: process.env.COHERE_API_KEY!,
  model: 'embed-english-v3.0',
})

Multilingual support

const embeddings = cohereEmbeddings({
  apiKey: process.env.COHERE_API_KEY!,
  model: 'embed-multilingual-v3.0',
})

Input Types

Cohere models support different input types optimized for specific use cases:

Type	Use Case
search_document	Embedding documents to be searched (default)
search_query	Embedding search queries
classification	Text classification tasks
clustering	Clustering similar texts

For memory storage and retrieval, search_document is typically the best choice.

Batch Processing

The provider automatically handles Cohere’s batch size limit (96 texts per request):

// Handled automatically - splits into batches of 96
await embeddings.embedBatch(texts) // Works with any size array