Skip to content

Reasoning Models (System 2)

New in v1.4.1

Lár treats "Reasoning Traces" (also known as Chain of Thought or System 2 Thinking) as a first-class citizen.

Instead of cluttering your final answer with <think> tags or internal monologues, Lár automatically extracts this thought process, saves it to a hidden metadata field for auditing, and delivers only the clean answer to your downstream nodes.

Supported Models

Lár supports reasoning capture for all major "Thinking" models:

Provider Model Method
DeepSeek deepseek-reasoner (R1) API Metadata (reasoning_content)
OpenAI o1-preview, o1-mini API Metadata (reasoning_content)
Ollama deepseek-r1, liquid-thinking Regex Fallback (<think> tags)
Liquid liquid-thinking Regex Fallback (<think> tags)

How it Works

When LLMNode executes, it performs a two-step extraction:

  1. API Check: It first checks if the API response contains a specific reasoning_content field (Standard for OpenAI/DeepSeek APIs).
  2. Regex Fallback: If not found, it scans the text for XML tags like <think>...</think>.
    • It extracts the content inside the tags.
    • It removes the tags and content from the output_key (the state).
    • It saves the extraction to run_metadata.

Robustness (v1.4.1)

The regex engine is robust to malformed outputs common in smaller local models: * Missing Closing Tag: Captures everything after <think> (e.g., if the model hit a token limit). * Missing Opening Tag: Captures everything before </think> (e.g., if the model started thinking "silently").

Usage

You do not need special configuration. Just use LLMNode as usual.

1. Using DeepSeek R1 (Ollama)

from lar import LLMNode

node = LLMNode(
    model_name="ollama/deepseek-r1:7b",
    prompt_template="Solve this riddle: {riddle}",
    output_key="answer"
)

The Result:

  • State (state['answer']): 
    The answer is a sponge.
  • Audit Log (run_metadata): 
    {
  "model": "ollama/deepseek-r1:7b",
  "total_tokens": 150,
  "reasoning_content": "First I need to analyze the properties... It holds water... It has holes..."
}

2. Using OpenAI o1

node = LLMNode(
    model_name="openai/o1-mini",
    prompt_template="Write a secure Python function for: {task}",
    output_key="code"
)

Best Practices

  1. Audit Logs: Always inspect lar_logs/ (or your custom log dir) to see the reasoning. It is the only place it is stored.
  2. Prompting: For local models (like Liquid Thinking), you may need to explicitly ask for tags in the system prompt if the model is stubborn: 
    system_instruction="You are a reasoning engine. You MUST enclose your internal monologue in <think> tags."