What is the difference between Command R and Command R+?

Both are tuned for retrieval-augmented generation and tool use. Command R+ is larger and more capable on complex reasoning and multi-step tools, while Command R is faster and cheaper for high-volume or simpler tasks. Start with R+ for quality, drop to R to cut cost.

Do I need the Embed and Rerank endpoints if I already use Chat?

For a basic chatbot, no. For RAG you typically use all three — Embed turns your documents and queries into vectors for retrieval, Rerank reorders the retrieved chunks by true relevance, and Chat generates the grounded answer. Rerank alone often gives the biggest accuracy jump.

What input_type should I pass to Embed?

Cohere's v3 embeddings are asymmetric, so you must tag the purpose. Use search_document when embedding the texts you store, and search_query when embedding the user's question at retrieval time. Mismatching them quietly degrades recall.

Is the API OpenAI-compatible?

No. Cohere uses its own v2 schema — a messages array for chat similar to OpenAI, but distinct embed and rerank endpoints with no OpenAI equivalent. Use the official Cohere SDK rather than pointing an OpenAI client at it.

How is Rerank priced and when is it worth it?

Rerank is billed per search (a query plus its candidate documents), separately from token-based chat pricing. It is worth it whenever your vector search returns a noisy top-k — reranking the top 50 down to the best 5 measurably improves answer quality for a tiny fraction of the chat cost.

Getting Started with the Cohere API

Why Cohere

Cohere is an enterprise-focused NLP provider whose models are built around retrieval and search rather than open-ended chat. Its three workhorse endpoints map cleanly onto the parts of a production RAG system: Embed turns text into vectors, Rerank scores how relevant documents are to a query, and Chat (Command R and R+) generates grounded answers with citations and tool calls. If you are building search, knowledge assistants, or document processing at scale, Cohere’s separation of these concerns is a genuine advantage over a single chat-only API.

Use the builder above to assemble a real request for any of the three endpoints, then copy the curl and run it with your own key.

How the three endpoints fit together

A typical Cohere RAG flow looks like this. At ingestion time you call Embed with input_type: "search_document" to vectorise every chunk of your corpus and store the vectors. At query time you embed the user’s question with input_type: "search_query", do an approximate-nearest-neighbour lookup to get a noisy top-k, then pass that candidate set to Rerank to reorder by true relevance. Finally you hand the best few chunks to Chat as documents so Command R+ produces an answer grounded in your data — with inline citations it generates automatically when you pass a documents field.

The key detail beginners miss is that v3 embeddings are asymmetric: documents and queries must be embedded with different input_type values or retrieval quality drops noticeably. Get that right and Cohere’s retrieval stack is hard to beat.

Tips and gotchas

Authenticate with a Bearer header. Cohere uses Authorization: Bearer $CO_API_KEY, not a custom key header.
Start with Rerank for cheap wins. If your RAG answers are mediocre, adding a Rerank pass over your existing retrieval is the highest-leverage change you can make and costs far less than the chat call.
Pin the model string. Always specify an exact model (e.g. command-r-plus) rather than an alias, so a model update never silently changes your behaviour.
Move to the SDK once it works. Prototype with the curl above, then switch to the official cohere Python or TypeScript SDK for retries, streaming, and typed responses in production.