兔子API 兔子API
中

provider

Cohere

Direct answerNew direct Cohere integrations use native POST /v2/chat. This site's /v1/chat/completions is an OpenAI-compatible route to an adapted channel, not Cohere's native path. Embed requires the correct input_type and dimension, while Rerank, Trial/Production limits, and model retirement must be checked separately.

Updated · Reviewed

Separate native Cohere V2 from this site's compatibility route

Cohere's native chat endpoint for new integrations is POST https://api.cohere.com/v2/chat. It uses a Cohere Bearer key, requires model, accepts chronological messages, and returns Cohere V2 content structures. This site's /v1/chat/completions accepts an OpenAI-compatible request and relays it through a configured Cohere channel. It is not an official Cohere URL, and this site does not expose /v2/chat here.

Embed input_type defines vector purpose

Embed V3 and newer require input_type. Use search_document for indexed corpus items and search_query for live queries; classification, clustering, and images have separate roles. Embed v4 and newer allow output_dimension values of 256, 512, 1024, or 1536, with 1536 as the default. A model, dimension, or embedding-type change requires rebuilding or parallel-migrating indexes. This page does not list an Embeddings route, so Cohere's /v2/embed must not be presented as a site endpoint.

Rerank is the second stage after retrieval

Rerank takes one query and candidate documents, returning original document index values and relevance_score values. Use it after keyword or vector retrieval to narrow a broad candidate set. Native Cohere V2 uses /v2/rerank; this site adapts /v1/rerank, and clients must follow each side's authentication and response contract. Long documents can be truncated or chunked, so evaluate ranking quality, latency, and search-unit cost.

Trial and Production limits vary by endpoint

Cohere does not publish one QPS for every capability. Its current table lists common Command Chat at 20 requests/minute for Trial and 500 for Production, Rerank at 10 and 1,000 requests/minute, and text Embed at 2,000 inputs/minute for both key types. Production access to some newer Chat models requires contacting sales, while Trial and some newer variants also carry monthly call limits. On 429, identify the exact model, endpoint, minute limit, or monthly limit before using bounded backoff.

Retirement makes old IDs fail

Cohere's Deprecations page lists retired models and endpoints by date with replacements. Embed v2.0 and selected Aya 8B IDs retired on 2026-04-04, and requests to them fail. Chat migration to V2 also combines history into messages and makes model required; Embed migration must validate embedding_types, input_type, dimensions, and vector indexes, not merely replace a model name.

Reviewed provider catalog

Latest models by capability

These are provider-catalog models, not a promise of availability on this site. Confirm callable IDs, endpoints, and pricing in the model marketplace.

Text, reasoning, and code

Speech and audio

Embeddings, retrieval, and reranking

Check live availability in the model marketplace

Use cases

  • Enterprise agents and RAG
  • Embeddings and reranking
  • Multilingual understanding and generation

API protocols

  • /v1/chat/completions
  • /v1/rerank

FAQ

Should a direct Cohere integration use /v2/chat or this site's /v1/chat/completions?

A new direct integration uses native POST https://api.cohere.com/v2/chat, a Cohere Bearer key, and the V2 messages and response schema. This site's /v1/chat/completions is an OpenAI-compatible ingress translated to a configured Cohere channel. It is not a native Cohere URL, and V2 response parsing cannot simply be copied to it.

How do I choose search_document versus search_query for Cohere Embed?

Use search_document for corpus items stored in a vector database and search_query for user queries searching that database. Classification and clustering use their matching types. Embed V3 and newer require input_type; choosing mismatched document and query roles can reduce retrieval quality.

Can I change the Cohere Embed output dimension at any time?

output_dimension is available only on Embed v4 and newer, with 256, 512, 1024, or 1536 dimensions and a 1536 default. Stored and query vectors must use the same model, dimension, and compatible normalization, so a change normally requires rebuilding or parallel-migrating the index.

Where should Cohere Rerank sit in a RAG pipeline?

First retrieve a broad candidate set with keyword or vector search, then send query and documents to Rerank and select top items using returned index and relevance_score. Cohere's native path is /v2/rerank; this site's adapted route is /v1/rerank. Do not mix their paths or response contracts.

How do Cohere Trial and Production rate limits differ?

Limits are model- and endpoint-specific, not one multiplier. The current official table lists common Command Chat at 20 requests/minute for Trial and 500 for Production, Rerank at 10 and 1,000 requests/minute, and text Embed at 2,000 inputs/minute for both. Newer models may require sales contact, and Trial or newer Chat variants can also have monthly call limits; always check the live table and account.

Can I migrate a retired Cohere model by changing only its alias?

Not safely. Requests to retired IDs fail; for example, Embed v2.0 and selected Aya 8B models retired on 2026-04-04. Follow the Deprecations replacement guidance and retest schemas, dimensions, indexes, quality, cost, and limits before production instead of changing only a string.

  • RAG, embedding, and reranking model guideRAG combines chunking, embeddings, retrieval, reranking, grounded generation, and evaluation; completeness comes from citable evidence and no-answer handling, not a larger model list.
  • Tool use and AI agent model guideAn agent is a controlled loop, not merely tool calling: structured schemas, least privilege, state, idempotency, timeouts, approval, audit, and replayable evaluation.
  • AI API quickstart: from API key to first requestAn AI API call needs a matching base URL, API key, model ID, and endpoint; start with minimal curl, then add error handling, cost controls, and key security.
  • Common AI API error codes and troubleshootingHandle AI API failures by parsing the HTTP status, provider code, error type, message, retry hints, and request ID together. Fix request, authentication, billing, and hard-quota causes first; retry only transient throttling, network failures, and selected 5xx errors with strict bounds.

Official sources

  1. Cohere Chat API v2 Official
  2. Cohere Embed API v2 Official
  3. Cohere Rerank API v2 Official
  4. Cohere Rerank Model Guide Official
  5. Cohere API Keys and Rate Limits Official
  6. Cohere Model and Endpoint Deprecations Official
  7. Migrate Cohere API v1 to v2 Official