The model is part of your infrastructure plan
Deploy’s company-server configuration separates the application from its model endpoint. The documented paths are Ollama and an OpenAI-compatible endpoint, including vLLM. Model choice still depends on the task, memory, context size and tool capabilities required.
Ollama configuration
In the company-server configuration, select the Ollama backend, set the model host URL and use an installed model tag. These are example values, not connection details for a public service.
MODEL_BACKEND=ollama
MODEL_URL=http://model-host.internal:11434
MODEL_ID=your-installed-model-tagOpenAI-compatible configuration
For a compatible server, use the /v1 base URL and the served model name. An OpenAI-compatible protocol does not mean the model is hosted by OpenAI.
MODEL_BACKEND=openai_compatible
MODEL_URL=https://model.company.example/v1
MODEL_ID=your-served-model-nameMake the endpoint reachable
The address must resolve and be reachable from the application container. localhost inside a container refers to that container, not a separate GPU server. Use your approved network address and transport configuration. Supply any endpoint credentials through the agreed secret configuration.
Separate the two connections
Your application or SDK connects to the Retvest API. The Deploy server connects to the configured model endpoint. A Retvest API key is not the model service’s credential; keep the two addresses and authentication settings distinct.
If an integration cannot connect, distinguish company API reachability from model reachability. An SDK request reaching the Retvest server does not establish that the model is available, loaded or capable of the requested workflow.
Choose around the work
Discuss context length, document size, expected simultaneous use and required tools. A compatible endpoint is a connection protocol, not a guarantee that every model supports every workflow. Keep the chosen model name and relevant configuration with the pilot’s operating record.