What Is Cohere and Why SpeedMVPs Uses It
Cohere was founded by former Google Brain researchers and has built their business around enterprise NLP rather than consumer-facing AI products. Command R+ is their flagship generative model, designed specifically for RAG and grounded generation: it has a 128k context window, strong citation generation (it can point to specific retrieved documents that support its claims), and low hallucination rates on knowledge retrieval tasks. Cohere's Embed API generates high-quality text embeddings for semantic search and vector similarity. Their Rerank API takes a query and a list of retrieved documents and scores them for relevance, significantly improving RAG pipeline accuracy over simple vector similarity search alone. These three components, Command R+, Embed, and Rerank, form a coherent stack for enterprise knowledge management products. SpeedMVPs uses Cohere on projects where the client has a large enterprise knowledge base, such as technical documentation, regulatory filings, or customer support content, and needs search accuracy that justifies the engineering investment in a proper retrieval pipeline. Cohere's Rerank API in particular is a high-value addition to any RAG product, often improving answer accuracy by 15-30% over a pure vector search baseline. Cohere also offers private cloud deployment options through their Enterprise tier, where models run in your AWS, Azure, or GCP environment with no data leaving your infrastructure. This is a strong option for financial services firms, healthcare organisations, and legal tech companies with strict data sovereignty requirements.
Setting Up Cohere in a Production AI Project
Install the Cohere SDK: pnpm add cohere-ai Basic chat generation: ```ts import { CohereClient } from 'cohere-ai' const cohere = new CohereClient({ token: process.env.COHERE_API_KEY }) const response = await cohere.chat({ model: 'command-r-plus', message: userMessage, documents: retrievedDocuments, // RAG documents }) ``` The documents parameter in the chat request is how Cohere implements grounded RAG. You pass retrieved text chunks as structured document objects, and Command R+ generates an answer that cites specific documents. The response includes a citations array showing which parts of the answer came from which documents, which is invaluable for products that need to show users their sources. For embedding: ```ts const embeddings = await cohere.embed({ texts: textChunks, model: 'embed-english-v3.0', inputType: 'search_document', }) ``` For reranking retrieved results: ```ts const reranked = await cohere.rerank({ model: 'rerank-english-v3.0', query: userQuery, documents: candidateDocuments, topN: 5, }) ``` The rerank call adds a second retrieval pass after initial vector search, re-scoring candidates by relevance to the query using a cross-encoder model. This step typically adds 100-200ms of latency but meaningfully improves answer quality and is worth the cost in accuracy-sensitive products. For production, token usage logging, retry logic, and rate limit handling follow the same patterns as other providers. Cohere's API rate limits are lower than OpenAI's on the trial tier, so production deployments should request increased limits from Cohere's enterprise team.
Key Features and Capabilities
Command R+ with grounded RAG is Cohere's strongest differentiator. The model is trained to generate answers that are anchored to provided documents, with explicit citation links. For products where answer accuracy and source transparency are critical, such as legal research tools, medical information products, or financial compliance assistants, this citation capability is a meaningful trust signal for end users. The Embed API supports both English and multilingual embedding models. The multilingual model handles 100+ languages with consistent vector space properties, meaning a query in French will retrieve relevant documents in English if the content matches. For multilingual enterprise knowledge bases, this is a significant practical advantage. The Rerank API is a standalone relevance scoring service. You can use it with any retrieval backend, including Pinecone, Weaviate, Supabase pgvector, or Elasticsearch. It takes a query and up to 1,000 candidate documents and returns relevance scores. Adding reranking to an existing RAG pipeline typically improves MRR (mean reciprocal rank) by a significant margin. Command R+ supports tool use for agent architectures, with a multi-step reasoning capability designed for research and analysis agents that need to call tools, reason about results, and call further tools before returning a final answer. Cohere's Data Privacy commitments are enterprise-grade. Their enterprise agreements include DPAs under GDPR, and their private deployment option eliminates API data exposure entirely. For FCA-regulated firms or NHS-adjacent products, the private deployment path is worth evaluating.
Real-World Workflow: Cohere in an AI MVP
A typical SpeedMVPs implementation using Cohere: a professional services firm needed an internal knowledge search product that allowed consultants to search across five years of project reports, client deliverables, and methodology documents stored in SharePoint. Accuracy was critical because consultants would use the search results to inform client-facing work, and confidently wrong answers were worse than no answer. The architecture used Cohere Embed to generate document embeddings during an initial ingestion batch, stored in Pinecone with document metadata. When a consultant queried the system, the query was embedded with Cohere Embed (inputType: 'search_query'), used to retrieve the top 20 candidates from Pinecone by vector similarity, then reranked with Cohere Rerank to the top 5. The top 5 documents were passed to Command R+ via the documents parameter, and the model generated a cited answer. The citation UI showed the answer alongside highlighted excerpts from source documents. Consultants could click through to the original SharePoint document. This transparency was important for a knowledge product where users needed to verify AI-generated summaries against source material. Recall@5 measured after reranking was substantially higher than pure vector search, validating the additional latency and cost of the rerank step. The client went from a simple keyword search to a RAG product in three weeks.
Cost and Pricing Considerations
Cohere's API pricing covers three products: generation (Command R+), embedding (Embed), and reranking (Rerank). Command R+ is priced per million tokens at rates competitive with mid-tier models from other providers. Embed is priced per million tokens of input text, which is very low cost for the value it provides. Rerank is priced per 1,000 searches, which adds up at high volume but is typically modest relative to the generation cost. For a typical RAG product with 10,000 queries per month, the combined Embed plus Rerank plus Command R+ cost is usually well within a range that justifies the accuracy improvement over cheaper alternatives. Build a usage model based on your expected query volume and average document count per retrieval. Cohere's enterprise tier includes volume discounts and private deployment options at negotiated rates. If your product expects significant scale, engage Cohere's enterprise sales team early to understand committed spend options. For GDPR purposes, Cohere's API processes data on their infrastructure. For products handling personal data, you need a DPA with Cohere. Their enterprise terms include this. The private deployment option eliminates the need for a third-party DPA entirely.
Alternatives to Cohere
For embedding and reranking specifically, Voyage AI (used by Anthropic) and Jina AI offer competitive embedding models. Jina is particularly strong for multilingual and long-document embeddings. For reranking, cross-encoder models from Hugging Face can be self-hosted as a cost-efficient alternative to Cohere Rerank at scale. For the generation component, OpenAI GPT-4o with a manually constructed citations prompt can approximate Command R+'s grounded generation, but requires more prompt engineering and produces less reliable citation accuracy. Claude Sonnet with documents passed in the context is another viable alternative. Pinecone and Weaviate both have native reranking integrations that can reduce the dependency on Cohere Rerank specifically. Weaviate's hybrid search combines BM25 and vector search in a single query, which can partially substitute for a separate reranking step. For teams who want everything from one provider and are already on OpenAI, the OpenAI Assistants API with file search implements a complete RAG pipeline without requiring Cohere. Quality is competitive for standard use cases but lacks Cohere's explicit citation tracking.