RequestAdjustment
What happened to one requested option that was not sent as asked (RFC 0008, #2115).
A client sends what the provider and model can take and reports the rest on Completion.adjustments, instead of dropping it silently. A caller that must not run without a particular option looks for that option in the list and treats its RequestAdjustment.Dropped as a failure. An option sent with another value is Adjusted, even when the caller left it at its CompletionOptions default, since a value is still sent. An option not sent at all is Dropped only when the caller set it away from that default: OpenAI leaves temperature, top_p and the penalties out of a reasoning model's request, and a default call reports none of them, so an explicit value equal to the default goes unreported.
Reported today: every option the org.llm4s.model.RequestTransformer drops or adjusts (Anthropic, OpenAI, Gemini, Vertex AI, Bedrock); a reasoning that Gemini, Vertex AI, Bedrock, Ollama or the standard OpenAI-compatible dialect does not send; the sampling options OpenAI leaves out of a reasoning model's request; and the penalties xAI leaves out of one. Other options a client leaves out, such as Anthropic's top_p and penalties, are not reported yet and follow under RFC 0008, so an empty list does not prove every option was sent.
reason, requested and sent are for people and logs; match on option and the case.
Attributes
- Source
- RequestAdjustment.scala
- Graph
-
- Supertypes
