Inference
Use language, vision-language, structured, vLLM, and pooling inference in Refiner
Refiner inference helpers wrap provider calls so model work can run inside pipelines with bounded concurrency, metrics, retries, and media handling.
| Page | Use it for |
|---|---|
| Generate Text | Text and chat-style inference inside transforms. |
| Multimodal and Structured Output | Images, videos, files, and validated JSON output. |
| Providers and vLLM | OpenAI-compatible, Google, Anthropic, and vLLM providers. |
| Pooling | Pooling/token-classification style inference used by reward scoring. |
For episode-level VLM workflows, see Subtask Annotation.