Cohere

Cohere builds AI models with a particular focus on enterprise search and retrieval. Their embedding models are designed specifically for search applications, with features like input type hints that let you tell the model whether you're embedding a document or a query. This can improve retrieval quality compared to general-purpose models.

Cohere's multilingual models are especially strong if you're working with content in multiple languages—they produce embeddings that work across languages, so a query in English can match documents in French or Japanese.

Setup

Install the Cohere SDK package:

bun add @ai-sdk/cohere

Set your API key in the environment:

COHERE_API_KEY="..."

Configure the provider in your unrag.config.ts:

import { defineUnragConfig } from "./lib/unrag/core";

export const unrag = defineUnragConfig({
  // ...
  embedding: {
    provider: "cohere",
    config: {
      model: "embed-english-v3.0",
      timeoutMs: 15_000,
    },
  },
} as const);

Configuration options

model specifies which Cohere embedding model to use. If not set, the provider checks the COHERE_EMBEDDING_MODEL environment variable, then falls back to embed-english-v3.0.

timeoutMs sets the request timeout in milliseconds.

inputType tells Cohere what kind of content you're embedding. This helps the model produce better embeddings for your specific use case.

truncate controls how the model handles input that exceeds its context window.

embedding: {
  provider: "cohere",
  config: {
    model: "embed-multilingual-v3.0",
    inputType: "search_document",
    truncate: "END",
    timeoutMs: 20_000,
  },
},

Input types

Cohere's embedding models support input type hints that optimize the embeddings for specific use cases:

search_document: Use this when embedding documents that will be searched. This is typically what you want for ingestion.

search_query: Use this when embedding search queries. The model optimizes for matching against documents.

classification: When embeddings will be used for classification tasks.

clustering: When embeddings will be used for clustering similar documents.

For RAG applications, you ideally want search_document when ingesting and search_query when retrieving. Unrag doesn't currently switch input types automatically, so the ingestion input type will be used for both.

Truncation options

NONE: Don't truncate. The API will return an error if the input is too long.

START: Truncate from the beginning of the input.

END: Truncate from the end of the input (most common choice).

Available models

embed-english-v3.0 is optimized for English text. It produces 1024-dimensional embeddings and is the best choice for English-only applications.

embed-multilingual-v3.0 supports over 100 languages and produces embeddings that work across languages. A query in one language can match documents in another.

embed-english-light-v3.0 and embed-multilingual-light-v3.0 are smaller, faster versions with slightly lower quality. Consider these if embedding speed is critical.

Environment variables

COHERE_API_KEY (required): Your Cohere API key. Get one from the Cohere dashboard.

COHERE_EMBEDDING_MODEL (optional): Overrides the model specified in code.

# .env
COHERE_API_KEY="..."
COHERE_EMBEDDING_MODEL="embed-multilingual-v3.0"

When to use Cohere

Choose Cohere when you're building search applications and want models designed specifically for retrieval, or when you need strong multilingual support. Their input type hints and search-focused optimization can give you an edge in retrieval quality.

Consider other providers when you're already invested in a different ecosystem or when you need multimodal embeddings (use Voyage instead).