AI provider setup
The AI assistant does not ship with a model. It talks to an API you provide, so before it can be switched on somebody has to create that API access on the provider's side and paste the result into Settings - General - AI assistant.
This page walks through creating those credentials for the two hosted providers, and covers the self-hosted option at the end.
Whichever provider you choose, SelectZero needs the same three things:
an API endpoint to send requests to
a model to run them against
an authentication token, unless the endpoint is open
OpenAI
You need an OpenAI account with billing enabled. API access is billed separately from a ChatGPT subscription, and a plain ChatGPT plan does not include it.
Sign in at platform.openai.com.
Open Settings → Billing and add a payment method. Without one the key authenticates but every request is refused for quota.
Go to API keys and choose Create new secret key. Give it a name that says where it is used, for example
selectzero.Copy the key immediately. OpenAI shows the value once and only stores a prefix afterwards, so if it is lost the key has to be replaced rather than re-read.
Decide which model to use. Any current chat model that supports structured (JSON) output will work; the assistant relies on that to return usable suggestions. Newer families are generally the better choice here, since suggestion quality tracks the model.
Then fill in the settings:
Field | Value |
|---|---|
API type |
|
API endpoint |
|
Model name | The model ID, for example |
Authentication type |
|
Authentication token | The secret key from step 4 |
Azure OpenAI
You need an Azure subscription with access to Azure OpenAI. Access is granted per subscription, so the resource type may not be available until that is approved.
In the Azure portal, create an Azure OpenAI resource. Pick the region deliberately: model availability differs by region, and the resource cannot be moved later.
Open the resource and go to Model deployments, then deploy the model you want. Give the deployment a name you will recognise, for example
selectzero-gpt5.Back on the resource, open Keys and Endpoint and copy the Endpoint and one of the two keys. Two keys exist so you can rotate one while the other stays live.
Then fill in the settings:
Field | Value |
|---|---|
API type |
|
API endpoint | The Endpoint from step 3 |
Service version | The API version to use, for example |
Model name | Your deployment name from step 2, not the model ID |
Authentication type |
|
Authentication token | A key from step 3 |
Ollama
Nothing has to be created with a vendor here, which is why this section is short: run the model on a host SelectZero can reach and point the endpoint at it. Pull the model you want on that host first, since the name below has to match a model it already serves.
Field | Value |
|---|---|
API type |
|
API endpoint | Your Ollama host, for example |
Model name | The pulled model tag, for example |
Authentication type |
|
The assistant only ever sends Ollama an endpoint and a model name, so the authentication token is ignored for this API type. Small models struggle with the structured output the assistant depends on, so anything below about 7B parameters is not worth using here.
Verifying the configuration
Use Test connection on the same tab. It sends a short prompt, so it exercises the endpoint, the model and the token together.
The button stays greyed out until Enable AI assistant is switched on, which is the first control on the tab, so turn that on before trying to verify.
It tests the values currently in the form, including edits you have not saved yet. Saving first is still worth doing, because the token field always starts empty and a blank token falls back to the stored key - so testing straight after a save is what confirms the combination that is actually stored.
A failure surfaces as a toast carrying the provider's own error. The usual causes are:
The endpoint is wrong, or the host is not reachable from the SelectZero deployment. Outbound access to the provider is a common omission on locked-down networks.
The token is wrong, revoked, or belongs to a different resource than the endpoint.
Endpoint, model name or authentication is missing.
The model is unknown - on OpenAI the model ID is wrong or not enabled for the account; on Azure the Model name is a model ID rather than the deployment name.
What leaves your deployment
Everything the assistant sends goes to the provider you configure, which for OpenAI and Azure OpenAI is a third party. Object and column names are always included. Three settings widen that, and all are described in General settings:
Include profiling values adds profiling statistics for the selected columns. On by default.
Include glossary terms adds the content of your business glossary terms. On by default.
Include sample data values adds real values from the data, and is the only one of the three that is off by default. Turning it on means actual row data leaves the deployment.
So on a fresh install profiling statistics and glossary content already go to the provider; only sample data has to be opted into.
A self-hosted Ollama endpoint keeps all of it inside your own network, which is usually the reason to choose one.