AI routing guide
Connect an OpenAI-compatible client to a configured provider.
Store a named provider credential, create the provider record, restart the server, then verify discovery, generation, streaming, errors, and usage.
- Provider record
- OpenAI-compatible
- Streaming
Connect credential and provider.
- Create an administrator-controlled credential containing the provider's required fields.
-
Create an enabled
llm_providerwith its type, comma-separated client-visible model names, optional base URL, andcredential_nameset to that exact name. - Confirm the named credential contains the fields required by that provider. Local providers may not require one.
- Restart the latest release so the runtime loads the provider.
-
List
/v1/models, then call the intended OpenAI-compatible endpoint with a Daptin token.
curl http://localhost:6336/v1/chat/completions \
-H "Authorization: Bearer $TOKEN" \
-H "Content-Type: application/json" \
-d '{"model":"client-visible-model","messages":[
{"role":"user","content":"Hello"}
]}'Test the contract your client uses.
- Model discovery contains only intended enabled models.
- Chat completions, legacy completions, embeddings, and tool calls are tested only where the provider supports them.
- Streaming produces correctly terminated compatible events and survives client cancellation.
- Invalid models, provider rejection, timeout, quota exhaustion, and malformed output remain distinguishable.
- Provider keys never reach browser or mobile clients.
- Model access permissions and Daptin authentication are tested with ordinary users.
- Token or compute usage reaches the intended meter when plans govern AI access.
Provider capability varies. An OpenAI-compatible route does not make every configured provider support every endpoint or parameter.
Provider matrix
Verify capabilities provider by provider.
After configuring providers, models, and server-held credentials, restart the latest release and test the exact routes the product will call.
- Confirm model listing and the intended provider/model mapping.
- Test normal and streaming responses separately.
- Exercise tools and embeddings only where that provider supports them.
- Confirm provider errors do not expose credentials.
- Do not build against Responses, image-generation, batch, or file endpoints; they are not exposed by the latest release.
- If usage is metered, verify token or compute accounting and exhausted-plan denial.