Cohere Embed 5 and Rerank 5

Cohere's search models are called through two LLMTR endpoints. The Embed 5 models turn text, images and inputs that combine both into vectors on /v1/embeddings. The Rerank 5 models order candidate documents against a query on /v1/rerank and, on request, return the most relevant passages inside each document. Both work with your existing LLMTR API key.

Model	Endpoint	Context	Inputs	Price ($/1M tokens)
cohere/embed-v5.0-pro	/v1/embeddings	128,000	Text, images, both together	0.12
cohere/embed-v5.0-fast	/v1/embeddings	128,000	Text, images, both together	0.08
cohere/rerank-v5.0-pro	/v1/rerank	32,768	Text	0.05
cohere/rerank-v5.0-fast	/v1/rerank	32,768	Text	0.02

Prices

Prices are US dollars per 1 million tokens and equal Cohere's list prices. The Embed 5 price in the table is for text tokens. Images are counted as tokens too and billed at $0.40 per 1M image tokens on both models; an image's token count grows with its resolution. The response's usage.prompt_tokens carries the sum of text and image tokens.

On the Rerank 5 models the billed tokens are query tokens x number of documents plus the tokens of all documents. The response's usage.total_tokens carries that number.

No platform margin is added to model prices; the margin applies only when you top up credit. See Billing for details.

Embed 5: first request

curl "$LLMTR_BASE_URL/v1/embeddings" \
  -H "Authorization: Bearer llmtr-your_key" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "cohere/embed-v5.0-pro",
    "input": ["The notice period for terminating the lease is thirty days."],
    "input_type": "document"
  }'

The response is OpenAI-shaped; the OpenAI SDK works with a changed base_url:

from openai import OpenAI

client = OpenAI(base_url="https://llmtr.com/v1", api_key="llmtr-your_key")

docs = client.embeddings.create(
    model="cohere/embed-v5.0-pro",
    input=["The notice period for terminating the lease is thirty days."],
    extra_body={"input_type": "document"},
)
print(len(docs.data[0].embedding))  # 2048

input_type

Queries and documents are embedded differently. For search, send documents with document and queries with query; sending both with the same type lowers accuracy.

input_type	When
document or search_document	Documents and chunks you add to the index. Used when the field is omitted.
query or search_query	The user's search query.
classification	Texts that will feed a classifier.
clustering	Texts to be clustered.

Any other value returns 400 invalid_request.

cohere/embed-v5.0-pro and cohere/embed-v5.0-fast share one vector space. You can index a collection with Pro and send queries with Fast; the vectors are comparable as long as you use the same dimensions value.

Images and mixed input

Images are sent as base64 data URIs (data:image/png;base64,...). PNG, JPEG, WebP and GIF are accepted. An image URL is not accepted and returns 400 invalid_request; read the image on your side and encode it as base64.

input can take two forms:

• A list of { "content": [...] } objects: each object produces its own vector.
• A single list of content parts ([{ "type": "text", ... }, { "type": "image_url", ... }]): all parts together produce one vector. Use this form to embed the text and the image of a PDF page together.

import base64
from openai import OpenAI

client = OpenAI(base_url="https://llmtr.com/v1", api_key="llmtr-your_key")

with open("product.png", "rb") as f:
    image = "data:image/png;base64," + base64.b64encode(f.read()).decode()

# Two vectors: one for the image alone, one for the text alone.
separate = client.embeddings.create(
    model="cohere/embed-v5.0-fast",
    input=[
        {"content": [{"type": "image_url", "image_url": {"url": image}}]},
        {"content": [{"type": "text", "text": "White t-shirt and blue jeans"}]},
    ],
    extra_body={"input_type": "document"},
)

# One vector: text and image together.
fused = client.embeddings.create(
    model="cohere/embed-v5.0-fast",
    input=[
        {"type": "text", "text": "Product card: white t-shirt, blue jeans"},
        {"type": "image_url", "image_url": {"url": image}},
    ],
    extra_body={"input_type": "document"},
)

Index images with document, and use query for the text you search them with.

Output size and type

Field	Description
dimensions	256, 512, 768, 1024, 1536 or 2048. 2048 when omitted. Any other value returns 400 invalid_request listing the accepted values. The output_dimension spelling is accepted too.
output_dtype	float (default), int8, uint8, binary or ubinary. binary and ubinary pack the vector into bits; the length is one eighth of the dimension. Cohere's single-value embedding_types array is accepted too.
encoding_format	float (default) or base64. base64 applies only to the float type and returns the float32 bytes of the vector; with the other types the vector comes back as a list of integers. The OpenAI SDKs send base64 on their own and read both forms.
truncation	true (default) shortens a long input from the end. With false, an input over 128,000 tokens is rejected with 400.

Embed 5 limits

• At most 96 inputs per request. More returns 400 invalid_request; split the list into batches of 96.
• At most 128,000 tokens per input.
• Nested arrays ([["chunk 1", "chunk 2"]]) are not accepted.

Rerank 5: request

curl "$LLMTR_BASE_URL/v1/rerank" \
  -H "Authorization: Bearer llmtr-your_key" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "cohere/rerank-v5.0-pro",
    "query": "How many days do I have to return a defective product?",
    "documents": [
      "Products can be returned within 14 days of delivery.",
      "Our stores are open on weekdays from 10:00 to 22:00.",
      "Defective products can be returned within 30 days and we cover the shipping cost."
    ],
    "top_n": 2,
    "relevant_spans": { "mode": "enabled", "max_spans": 2 }
  }'

Field	Required	Description
query	Yes	The query the documents are scored against.
documents	Yes	Candidate documents, at most 1000 items. Write records such as tables or JSON as YAML text.
top_n	No	Returns only the N highest-scoring results. The top_k spelling is accepted too.
return_documents	No	When true, each result also carries its document text.
max_tokens_per_doc	No	The most tokens read from one document. When omitted a document is read up to the model's context; anything longer is shortened.
relevant_spans	No	Context Focus. mode is required (enabled or disabled). threshold is between 0 and 1, default 0.1; only passages scoring above it are returned. max_spans is the most passages per document; -1 returns all of them.

Rerank 5 shortens a long document instead of rejecting it. truncation: false is therefore not accepted and returns 400 invalid_request; set the limit with max_tokens_per_doc.

relevant_spans and max_tokens_per_doc are valid on these two models only. Sent to another rerank model, the request is rejected with 400 invalid_request.

Context Focus response

{
  "object": "list",
  "data": [
    {
      "index": 2,
      "relevance_score": 0.3949,
      "relevant_spans": [{ "start": 0, "end": 79, "score": 0.9950 }]
    },
    {
      "index": 0,
      "relevance_score": 0.3448,
      "relevant_spans": [{ "start": 0, "end": 65, "score": 0.9935 }]
    }
  ],
  "model": "cohere/rerank-v5.0-pro",
  "usage": { "total_tokens": 132 }
}

start and end are the start and end of the passage inside the document, counted in Unicode code points; end is exclusive. In Python documents[i][start:end] works directly. JavaScript strings are UTF-16, so split the text into code points with Array.from(text) first:

const chars = Array.from(documents[item.index]);
const passage = chars.slice(span.start, span.end).join("");

Sending these passages to the language model instead of whole documents keeps the context and the token spend smaller.

Scores and thresholds

relevance_score is between 0 and 1. Rerank 5 scores are calibrated across queries and domains: one threshold can drop low-scoring results across different queries. Pick the threshold with a few sample queries on your own data.

Choosing a model

• cohere/embed-v5.0-pro: highest accuracy; for indexing a collection and for searches where accuracy decides.
• cohere/embed-v5.0-fast: lower latency and price; for the query side, agent loops and high-volume search.
• cohere/rerank-v5.0-pro: complex queries and long or domain-specific documents.
• cohere/rerank-v5.0-fast: high-volume endpoints that rerank every query.

Errors

Status	Meaning
400 invalid_request	Invalid field: more than 96 inputs, an image URL, an unsupported dimensions, input_type or output_dtype, rerank with truncation: false, an out-of-range relevant_spans value. The message names the field.
404 model_not_found	The model ID is not in the catalog.
429 rate_limit_error	The API key's request limit was exceeded.
502 provider_error	The provider could not complete the request. Try again later.
503 model_unavailable	The model is temporarily unavailable. Try again later.

For the general request shapes see Embeddings and Rerank.