Endpoints
List past inferences
GET /inferences returns a paginated list of past inference calls. Use the query parameters below to filter results.
Query parameters
number
Maximum number of results to return per page.
number
Number of results to skip before returning. Use with
limit to paginate through results.string
Filter by model ID. Accepts a training job ID or a base model ID.
string
Filter by task type (e.g.
ner, classification, generate).string
Filter by project ID to see only inferences scoped to a specific project.
string
Filter by training job ID to see only inferences run against a specific fine-tuned model.
number
Minimum end-to-end latency in milliseconds (inclusive). Must be >= 0.
number
Maximum end-to-end latency in milliseconds (inclusive). Must be >= 0 and >=
latency_min if both are set.number
Minimum LLM-as-Judge score (inclusive), in the range
0.0–1.0.number
Maximum LLM-as-Judge score (inclusive), in the range
0.0–1.0. Must be >= llmaj_score_min if both are set.string
Inclusive lower bound on
created_at, as an ISO 8601 UTC timestamp.string
Exclusive upper bound on
created_at, as an ISO 8601 UTC timestamp.latency_min/latency_max and llmaj_score_min/llmaj_score_max each return 422 if the min value is greater than the paired max value.Example
Get inference details
GET /inferences/:id returns the full record for a single past inference, including the input text, schema, model response, and timestamp.
Submit feedback
POST /inferences/:id/feedback lets you mark a past inference as correct or incorrect, optionally attaching the corrected output. Incorrect verdicts with a correction are used as labeled training examples for Adaptive Inference.
Feedback submitted here powers Adaptive Inference — Pioneer’s continuous improvement loop that automatically retrains your model on corrections collected from live traffic. See the Adaptive Inference guide for details on how this works.
Request parameters
string
required
Human judgment on the inference:
correct or incorrect.object
The expected output, in the same shape as the original inference’s output (for example, a corrected
entities list for an NER inference). Required when verdict is incorrect; must be omitted or null when verdict is correct. Sending an incorrect combination returns 422.string
Optional free-text reviewer notes. Maximum 5000 characters.
Example
string
The inference that was annotated.
string
The stored verdict.
string
ISO 8601 timestamp of when the feedback was submitted.
Get feedback
GET /inferences/:id/feedback returns the feedback previously submitted for a specific inference. Returns 404 if no feedback has been submitted yet.
Related
- Pioneer native inference — run new inferences
- Adaptive Inference guide — continuous model improvement from live traffic