Skip to main content
The Pioneer inference endpoint accepts a model ID, input text, and a schema that defines exactly what to extract. You can target a fine-tuned model from a completed training job or call a base model directly. For encoder models (GLiNER), use the schema field to declare entities, classifications, structures, or relations. For decoder models, use "task": "generate" instead.

​
Endpoints

​
List the model catalog

Use GET /base-models to fetch the current list of available models. Filter by ?supports_inference=true to narrow to inference-ready models, and by ?task_type=encoder or ?task_type=decoder to filter by architecture.

​
Run inference

​
Request parameters

​
string
required
The ID of the model to run inference against. Use the job ID returned by POST /felix/training-jobs (e.g. job_abc123) to target a fine-tuned model, or a base model ID like fastino/gliner2-base-v1 to call a base model directly.
​
string | string[]
required
The input text to run the model against. Pass an array of strings to run batch inference — the response result
​
string[] | object
Defines what to extract from the input text. Used with encoder models. For simple NER, pass a flat array of entity labels (e.g. ["organization", "product"]). For multi-task extraction, pass an object with any combination of the following keys:
​
string[]
List of entity type labels to extract (Named Entity Recognition). Example: ["organization", "product", "location"].
​
object[]
List of classification tasks. Each object has a task string (the classification label group name) and a labels array of candidate class strings.
​
object
Dictionary of structure definitions for JSON extraction. Each key is a structure name; the value defines the shape of the output.
​
object[]
List of relation definitions. Each object describes a directional relationship between entity types to extract.
​
number
default:"0.5"
Confidence threshold for returned predictions. Values range from 0 to 1. Lower values return more candidates at the cost of precision; higher values return fewer, higher-confidence results.

​
Example — NER with a fine-tuned model

​
Example — combined schema (entities + classifications)

Decoder models use a different request shape. Pass "task": "generate" and a messages array instead of text and schema:
The threshold default is 0.5. Lower it (e.g. 0.3) to surface more candidates at the cost of more false positives. Raise it (e.g. 0.7) for tighter, higher-precision extractions.

​
Using a base model ID

If you haven’t fine-tuned a model yet, you can call a base model directly. Use a model ID from GET /base-models, such as fastino/gliner2-base-v1.