The Pulse
Cohere Makes Embed 5 and Rerank 5 Available on OpenRouter
Cohere says Embed 5 and Rerank 5 are now available on OpenRouter through a single API. Its walkthrough shows how the models can be used in a retrieval-augmented generation pipeline that searches documents and supplies selected text to a cha
AI.info Team · Written with AI · fact-checked ·
Cohere says Embed 5 and Rerank 5 are now live on OpenRouter, with access through a single API. Its demonstration builds a retrieval-augmented generation (RAG) pipeline, which retrieves relevant text from a document collection before a language model answers a question.
One OpenRouter key for retrieval and generation
OpenRouter users can use the same API key and credits they already use for other models, with no additional setup. Cohere’s walkthrough uses one client for embedding documents, retrieving candidates, reranking them and sending selected text to a chat model. The workflow needs no separate SDKs, accounts or bills to manage.
The pipeline has an offline indexing stage and an online path that runs for each question. Indexing embeds document chunks to make them searchable; each question then calls an Embed 5 Fast model, a Rerank 5 model and a chat model through OpenRouter’s API. Cohere frames knowing which calls happen once and which happen per request as a way to explain most of the pipeline’s cost and latency.
Embed 5 splits indexing from query work
The example indexes documents with Embed 5 Pro and handles questions with Embed 5 Fast. Cohere says the two models share an embedding space, allowing Fast to search an index built with Pro as long as both use the same vector dimensions. In Cohere’s tests across 40 datasets, Fast queries on a Pro index kept 98.4% of all-Pro retrieval quality; Fast on both sides kept 96.6%.
Embed 5 produces vectors at six sizes, with 2,048 dimensions as the default. Cohere recommends 1,024-dimensional int8 vectors as the best quality-to-cost point, while noting that OpenRouter’s embeddings API returns floats; the post’s sample therefore uses 1,024-dimensional floats. The walkthrough keeps its vector store in NumPy for simplicity and says it can be replaced in production with a preferred vector database such as pgvector, Qdrant, Pinecone or Weaviate.
Rerank 5 filters the shortlist
The walkthrough’s retrieval function requests up to 20 chunks by cosine similarity. Its sample corpus contains six documents, each producing one chunk, so the example can return at most six candidates. Rerank 5 then re-scores the candidates, returns up to five, and the sample code drops any result with a relevance score below 0.3. Fewer than five results may therefore remain. Cohere describes Rerank 5’s scores as calibrated across queries and domains, so the same relevance threshold can filter weak context before it reaches the language model.
Rerank 5 can score a query and chunk together with up to 32,000 tokens combined; the walkthrough says longer inputs are split into several passes. Cohere says it reranks semi-structured material such as JSON records and tables well when formatted as YAML strings, and that switching rerank models does not require re-indexing because reranking reads raw text at query time. The post recommends Rerank 5 Pro for reasoning-heavy queries and long or domain-specific documents, and Rerank 5 Fast for high-volume search and agents that search many times per task.
A cited answer comes after retrieval
Once reranking selects the chunks, the sample sends them as numbered sources to a chat model and asks it to answer from those sources alone. The example uses deepseek/deepseek-v4.1-flash-1017, while its code comment says any chat model on OpenRouter can fill that role. If no result clears the relevance threshold, the sample skips the language-model call.
The sample compares the rankings for a question about rotating an API key without downtime. Embed 5 alone puts Webhook signing secrets second at 0.337 and Creating API keys third at 0.316; after reranking, both fall below the 0.3 cutoff. Rotating API keys scores 0.522 in the embedding-only results and 0.808 after Rerank 5.