Options for LLM API calls. More...

#include <llm_interface.h>

Public Attributes
double	temperature = 0.7
	The temperature of the LLM.
int	max_tokens = 1024
	The maximum number of tokens.
double	top_p = 1.0
	The top p of the LLM.
double	presence_penalty = 0.0
	The presence penalty of the LLM.
double	frequency_penalty = 0.0
	The frequency penalty of the LLM.
int	timeout_ms = 30000
	The timeout in milliseconds.
std::vector< std::string >	stop_sequences
	The stop sequences of the LLM.
std::optional< JsonObject >	response_schema
	Response schema for structured output (JSON schema).
std::optional< std::string >	response_mime_type
	Response MIME type for structured output.
std::optional< std::string >	reasoning_effort
	Reasoning effort for thinking-capable models.

Detailed Description

Options for LLM API calls.

Examples: autonomous_agent_example.cpp, combined_workflows_example.cpp, coroutine_example.cpp, evaluator_optimizer_example.cpp, multimodal_example.cpp, orchestrator_example.cpp, parallel_example.cpp, prompt_chain_example.cpp, robotics_object_detection_demo.cpp, routing_example.cpp, simple_agent.cpp, and streaming_chat.cpp.

Member Data Documentation

std::optional<std::string> agents::LLMOptions::reasoning_effort

Reasoning effort for thinking-capable models.

Accepted values: "low", "medium", "high". Mapped per provider:

Ollama (gpt-oss): sets think to the level string (Reasoning: low|medium|high in template).
OpenAI (o1/o3/o4-mini/gpt-5): sets reasoning_effort API field.
Anthropic (Claude 4.x with extended thinking): sets thinking.budget_tokens using a level→budget mapping (low=1024, medium=4096, high=16384).
Google (Gemini): no equivalent yet; ignored.

If unset (nullopt): Ollama defaults to think: true (model default level + channel split); other providers don't send any reasoning param.

std::optional<std::string> agents::LLMOptions::response_mime_type

Response MIME type for structured output.