org.llm4s.llmconnect.model.Thinking
See theThinking companion object
@Stable
Whether the model thinks before it answers, and how its thinking is bounded: separate from ReasoningEffort, which says how hard to think.
Set it with CompletionOptions.withThinking. A provider maps each case to its own control, and reports what it could not send as asked on Completion.adjustments (RFC 0008):
- Thinking.Off: do not think. Anthropic
thinking: disabled; on OpenAI, the lowest effort the model accepts. - Thinking.Adaptive: the model decides whether and how much to think. Anthropic
thinking: adaptive. - Thinking.Budget: think within a token budget. Anthropic
budget_tokens, fitted to the provider's range and reported asAdjustedwhen it does not fit.
@Stable, not @Experimental as RFC 0008 first wrote it: the @Stable CompletionOptions exposes it, as RequestOption is exposed by Completion (RFC 0008 decisions 12 and 14).
Attributes
- Example
-
CompletionOptions().withThinking(Thinking.Adaptive).withReasoning(ReasoningEffort.Low) Thinking.budget(8000).map(CompletionOptions().withThinking) // Result[CompletionOptions] - Companion
- object
- Source
- Thinking.scala
- Graph
-
- Supertypes
- Known subtypes
In this article
