AI and agents

How to use Cohere with DataNature: sovereign RAG and agents on your own data

DataNature connects to Cohere’s Command, Embed and Rerank models through Cohere’s API, its OpenAI-compatible Compatibility API, a dedicated Model Vault, or a private deployment inside your own VPC or data centre. Lakehouse tables and documents stay governed in DataNature; Cohere models generate, embed and rerank; Atlas logs every prompt and answer.

Published . Last updated

On this page

Why pair Cohere with DataNature?

Cohere builds enterprise models and offers deployment choices most model providers do not: SaaS, a dedicated Model Vault, a public or hybrid cloud VPC, and fully private deployment on-premises. DataNature is the governed data layer those models need. Together they give a Canadian-built stack for retrieval-augmented generation (RAG) and agents, with data and AI kept in one jurisdiction.

Which Cohere models fit which job?

JobCohere modelWhere it runs in the flow
Answers, reasoning and tool useCommand family (Command A+)Voir agents and AI360 Studio decisions
Semantic search over documents and tablesEmbedVectors stored in DataNature AI's vector database
Better top results for RAGRerankBetween retrieval and generation
Agentic codingNorth Mini CodeDeveloper tools, data engineering assistants
Workplace assistant and agent builderNorthAlongside Voir, on governed data

What are the integration options?

  1. Cohere API. Fastest start. Good for non-sensitive data or proofs of concept.
  2. Compatibility API. Cohere exposes an OpenAI-compatible endpoint, so tools built for that interface can switch with a base URL and key.
  3. Model Vault. Dedicated Cohere capacity hosted by Cohere.
  4. Private deployment. Cohere model containers run in your VPC or on your own GPUs, typically on Kubernetes, next to DataNature. Nothing leaves your boundary.

For regulated banks, credit unions and governments, option 4 inside a Canadian cloud such as R2i Edwin keeps data, prompts and outputs in Canada.

What does the architecture look like?

  1. DocuContext parses PDFs, scans and emails into clean text and fields.
  2. Embed turns chunks into vectors stored in DataNature AI, with Atlas permissions copied from the source.
  3. A question arrives in Voir. Retrieval pulls candidate chunks and lakehouse rows the user is allowed to see.
  4. Rerank orders them; Command writes the answer with citations.
  5. Atlas AI gateway records the prompt, sources, model and cost.

How do permissions carry through RAG?

Retrieval filters by the user's Atlas permissions before anything reaches the model, so an agent never sees a document its user could not open. Row and column rules apply to tables the same way.

Can we mix Cohere with open models?

Yes. The Atlas AI gateway routes each use case to an approved model: Cohere for multilingual answers and reranking, an open-weight model for code, a small local model for classification. See three ways to run LLMs.

Sources

How is Cohere connected, technically?

Models are reached through the DataNature AI gateway, which an admin points at an approved endpoint. Until it is configured it shows "Not configured" and agents cannot call any model. For Cohere, that endpoint can be a Cohere deployment in your private cloud or on-premises, or Cohere's API where policy allows. Typical routing:

TaskWhere it runs
Embeddings for documentsDocuContext generates embeddings into Milvus, inside the deployment
Reranking retrieved passagesCohere Rerank through the gateway
Answer generationCohere Command or another approved model through the gateway
Numeric questionsSemantic-layer cubes over Gold tables, not the model

Because routing is per endpoint, a bank can send its most sensitive twins to an in-country model and use a provider endpoint only for non-sensitive work. See private RAG on governed lakehouse data.

See it on your own data.

Book a demo

Demos and quotes are handled by Dlytica, the company behind DataNature.