Model discovery and probe objects¶
Before you call this API¶
Use BASE_URL, MOI_KEY, and WORKSPACE_ID from the same target environment. Obtain BACKEND_ID from its backend list. Verify production deployment support separately. Provider addresses and secrets are placeholders; replace them before actual calls. These product APIs manage model supply separately from Genesis inference/administration APIs.
Root /llm, /embedding, and /parser paths still select tenants through X-Workspace-ID. No workspace segment does not imply global administrator configuration. Root endpoints use regular product identity. Listed workspace-path endpoints additionally accept verified task identity.
Model lists filter by model_backend.read. Saved-backend probes require read on that object. Temporary-address probes use regular access and cannot read unauthorized saved configuration. Configuration writes separately require create/update/delete and router authorization. This does not imply model runtime privileges.
Response fields¶
Field |
Type |
Description |
|---|---|---|
data.models |
array (array of objects or strings) |
Lists return objects; probes return model-name strings. |
models[].model |
string |
Model name. |
models[].backend_id |
integer |
Backend ID owning the routed model, possibly a synthetic system ID. Do not use negative IDs with management APIs requiring positive values. |
models[].backend_name |
string |
Backend name. |
models[].model_type |
string |
Type marker in the root model list. |
models[].reasoning_efforts |
array (array of strings) |
Available reasoning levels; omitted when empty. Regular workspace declarations do not guarantee this field. |
models[].source |
string |
workspace/system source in the root list. |
models[].system_default |
boolean |
System-default marker in the root list. |
models[].type |
string |
Optional embedding subtype, not a backend protocol such as OPENAI. |
models[].dim |
integer |
Optional embedding dimension; omitted when zero. |
data.backend |
object |
Backend saved by the set operation. |
data.endpoint |
object |
Endpoint saved by the same set operation. |
Workspace embeddings/models projects only model/backend_id/backend_name plus nonempty type/nonzero dim and returns code=0. Root lists use string OK/msg=OK. Workspace llm/models projects names, backends, and request-type markers only, without merging system lists. These three protocols are not aliases. None is paginated.
Backend/endpoint field |
Type |
Description |
|---|---|---|
backend.id |
integer |
Backend ID. |
backend.name |
string |
Name. |
backend.type |
string |
Backend protocol type. |
backend.reasoning_control_protocol |
string |
Reasoning control protocol. |
backend.timeout_seconds |
integer |
Configured timeout in seconds. |
backend.models |
array (array of strings) |
Declared model names. |
backend.endpoints |
array (array of objects) |
May include nonempty endpoints; set results separately provide data.endpoint. |
backend.created_at |
integer |
Creation time in Unix seconds. |
backend.updated_at |
integer |
Update time in Unix seconds. |
endpoint.id |
integer |
Endpoint ID. |
endpoint.backend_id |
integer |
Owning backend ID. |
endpoint.address |
string |
Access URL. |
endpoint.status |
string |
ONLINE/OFFLINE, separate from numeric 0/1 in workspace Core endpoints. |
endpoint.created_at |
integer |
Creation time in Unix seconds. |
endpoint.updated_at |
integer |
Update time in Unix seconds. |
Credentials are absent from these responses. Reasoning control supports NONE, OPENAI_REASONING_EFFORT, CHAT_TEMPLATE_KWARGS_ENABLE_THINKING, CHAT_TEMPLATE_KWARGS_THINKING, OLLAMA_THINK, QIANWEN_ENABLE_THINKING, and corresponding formal REASONING_CONTROL_* values. The actual model must support the chosen protocol. Saving performs no capability test.
Probe results and side effects¶
Probes append /models to the specified root URL, read provider data[].id, remove empty names, and classify LLM/Embedding by naming patterns. They do not test vision, OCR, reasoning, or actual embeddings. Saved-backend probes require explicit api_key. LLM DAO filters offline endpoints before selecting the first online candidate. Embedding retains both states and takes the first address without rotation. Candidates are not written back. Probes read at most1 MiB; larger responses fail. Provider networking, timeouts, rate limits, and authentication require live verification.
Declared source/type/dim values do not replace actual model calls. Root-list service_slot additionally filters by registered configuration; the two workspace model lists ignore this parameter. Binding models, changing routes, and inference are separate subsequent tasks.
States and errors¶
Legacy root endpoints use string code/msg. Parameter/workspace errors return 400, missing objects404, and denied privileges403. Probe failures return 502/ErrProbe; system/authorization dependency failures have other mappings. Workspace embedding reads use numeric code/message. Root binding decodes one JSON value without common strict unknown-field rejection. Resource-tag binding rules do not apply here.