Reasoning Effort Control

Some LLM models support extended thinking modes that allow them to reason more deeply about complex problems before responding. Mojentic provides reasoning effort control through the CompletionConfig.

Overview

The reasoning_effort field allows you to control how much computational effort the model should spend on reasoning:

• Low: Quick responses with minimal reasoning
• Medium: Balanced approach (default reasoning depth)
• High: Extended thinking for complex problems

Usage

use mojentic::llm::gateway::{CompletionConfig, ReasoningEffort};
use mojentic::llm::broker::LlmBroker;
use mojentic::llm::models::LlmMessage;
use mojentic::llm::gateways::ollama::OllamaGateway;
use std::sync::Arc;

#[tokio::main]
async fn main() -> anyhow::Result<()> {
    let gateway = Arc::new(OllamaGateway::new());
    let broker = LlmBroker::new("qwen3:32b", gateway, None);

    // Configure with high reasoning effort
    let config = CompletionConfig {
        reasoning_effort: Some(ReasoningEffort::High),
        ..Default::default()
    };

    let messages = vec![
        LlmMessage::system("You are a helpful assistant."),
        LlmMessage::user("Explain quantum entanglement in simple terms."),
    ];

    let response = broker.generate(&messages, None, Some(config), None).await?;
    println!("Response: {}", response);

    Ok(())
}

Provider-Specific Behavior

Ollama

When reasoning_effort is set (any value), Ollama receives the think: true parameter, which enables extended thinking mode. The model’s internal reasoning is captured in the thinking field of the response:

#![allow(unused)]
fn main() {
let response = broker.generate(&messages, None, Some(config), None).await?;

// For Ollama, check the gateway response directly for thinking field
// The thinking content shows the model's internal reasoning process
}

OpenAI

For OpenAI reasoning models (like o1), the reasoning_effort parameter maps directly to the API’s reasoning_effort field:

• ReasoningEffort::Low → "low"
• ReasoningEffort::Medium → "medium"
• ReasoningEffort::High → "high"

For non-reasoning OpenAI models, the parameter is ignored and a warning is logged.

When to Use Reasoning Effort

Use higher reasoning effort for:

• Complex mathematical problems
• Multi-step logical reasoning
• Code generation requiring architectural decisions
• Nuanced ethical or philosophical questions
• Tasks requiring careful consideration of trade-offs

Use lower reasoning effort for:

• Simple factual questions
• Quick classifications
• Routine formatting tasks
• When response speed is critical

Performance Considerations

Higher reasoning effort typically means:

• Longer response times
• Increased token usage
• More thoughtful, accurate responses
• Better handling of edge cases

The trade-off between speed and quality should guide your choice of reasoning effort level.