Cohere Embeddings (via AWS Bedrock)

When using Cohere embeddings, llm-exe will make POST requests to the AWS Bedrock endpoint for your configured region. Both Cohere Embed v3 and Embed v4 are supported through a single provider key — pick the variant via the model option.

Options

Option	Type	Default	Description
model	string	—	The Bedrock model ID (e.g., cohere.embed-english-v3, cohere.embed-multilingual-v3, cohere.embed-v4:0)
inputType	"search_document" | "search_query" | "classification" | "clustering"	"search_document"	How Cohere should prepare the embedding. Use search_document for the corpus you index and search_query for queries you embed at search time
truncate	"NONE" | "START" | "END" | "LEFT" | "RIGHT"	—	How over-length inputs are handled. START/END for v3, LEFT/RIGHT for v4. NONE returns an error if the input is too long
dimensions	number	—	Output vector size. For Embed v4: 256, 512, 1024, or 1536 (maps to Cohere's output_dimension). For Embed v3: the model has a fixed 1024-dim output — passing 1024 is accepted as a no-op, any other value throws immediately so you don't silently get a different dimension than you asked for
awsRegion	string	AWS_REGION env var	The AWS region for the Bedrock endpoint (required)
awsSecretKey	string	—	AWS secret key (if not using default credentials)
awsAccessKey	string	—	AWS access key (if not using default credentials)

Basic Usage

import { createEmbedding } from "llm-exe";

const embeddings = createEmbedding("amazon:cohere.embedding.v1", {
  model: "cohere.embed-multilingual-v3",
  awsRegion: "us-west-2",
});

const embedding = await embeddings.call("The string of text you would like as vector");
const vector = embedding.getEmbedding();
console.log(vector);
// [
//   -0.0419921875, 0.005367279052734375, 0.011688232421875,
//   -0.0096893310546875, 0.039459228515625, -0.0231781005859375,
//   // ...
// ]

Batching

Cohere accepts up to 96 texts per call. Pass an array to embed them in a single request:

const embedding = await embeddings.call([
  "RAG system design patterns",
  "Actuarial loss triangles and reserving primer",
]);

// One vector per input text
const first = embedding.getEmbedding(0);
const second = embedding.getEmbedding(1);

Search vs Indexing

For retrieval (RAG) workflows, embed your corpus with inputType: "search_document" and your queries with inputType: "search_query". Cohere prepends different special tokens for each case, which improves retrieval quality.

// At index time
const docEmbeds = createEmbedding("amazon:cohere.embedding.v1", {
  model: "cohere.embed-english-v3",
  inputType: "search_document",
});

// At query time
const queryEmbeds = createEmbedding("amazon:cohere.embedding.v1", {
  model: "cohere.embed-english-v3",
  inputType: "search_query",
});

Embed v4 Output Dimensions

Cohere Embed v4 supports variable output dimensions. Pass dimensions to pick one of 256, 512, 1024, or 1536 (default 1536 if unset). Smaller vectors trade off some retrieval quality for cheaper storage and faster nearest-neighbor lookups.

const embeddings = createEmbedding("amazon:cohere.embedding.v1", {
  model: "cohere.embed-v4:0",
  dimensions: 512,
});

Multimodal Input (Embed v4)

Cohere Embed v4 can embed text and images together. Pass an array of EmbeddingContentItem objects instead of strings, and llm-exe routes them to Cohere's inputs field for you:

import { createEmbedding } from "llm-exe";
import type { EmbeddingContentItem } from "llm-exe";

const embeddings = createEmbedding("amazon:cohere.embedding.v1", {
  model: "cohere.embed-v4:0",
  awsRegion: "us-west-2",
});

const items: EmbeddingContentItem[] = [
  {
    content: [
      { type: "text", text: "A diagram of the retrieval pipeline" },
      { type: "image_url", image_url: { url: "data:image/png;base64,iVBORw0KGgo..." } },
    ],
  },
];

const embedding = await embeddings.call(items);
const vector = embedding.getEmbedding(0);

Each EmbeddingContentItem collapses into one vector. That is the point: a caption and its image become a single searchable embedding, rather than two vectors you have to reconcile later.

Things to know:

• image_url.url must be a complete data URI (e.g. data:image/png;base64,...). Bedrock does not fetch remote http(s) URLs — read the bytes yourself and inline them.
• Use a multimodal model. Embed v3 has no image support, and Bedrock will reject the request.
• A batch must be uniform. Mixing strings and content items in one call throws embedding.unsupported_input, because Cohere's texts and inputs fields are mutually exclusive and silently splitting the batch would misalign every downstream index. Embed text and multimodal entries in separate calls.
• Only amazon:cohere.embedding.v1 accepts multimodal input. Passing content items to openai.embedding.v1 or amazon.embedding.v1 throws embedding.unsupported_input up front instead of producing an opaque provider-side 400.

Notes

• Cohere on Bedrock does not include token usage in the response body. llm-exe reads it from Bedrock's x-amzn-bedrock-input-token-count / x-amzn-bedrock-output-token-count response headers, so getResult().usage.input_tokens is populated when those headers are present and falls back to 0 only when they are absent.
• Cohere models on Bedrock are served from AWS Marketplace. The first invocation in your account must come from a user with aws-marketplace:Subscribe permissions, after which all users in the account can invoke the model.
• CohereBedrockEmbeddingOptions also exposes an imageInputs option that writes Cohere's inputs field directly. It is an escape hatch — prefer passing content items as the call input, which takes precedence over the option.