agentchat.contrib.reasoning_agent

​

ThinkNode

class ThinkNode()

​

__init__

def __init__(content: str, parent: Optional["ThinkNode"] = None) -> None

A node in a tree structure representing a step in the reasoning process.

This class implements a tree node that stores content (text describing a reasoning step), maintains parent-child relationships, tracks node statistics, and provides utilities for traversing/visualizing the reasoning path.

Arguments:

• content str - The text content/description for this reasoning step.
• parent Optional[ThinkNode] - The parent node in the tree, if any.

Attributes:

• content str - The text content/description for this reasoning step.

• value Optional[float] - A numeric score/value assigned to this node.

• parent Optional[ThinkNode] - Reference to the parent node.

• reflection str - A string containing reflections on the reasoning process.

• rating_details str - A string providing details about the rating of this node.

• depth int - The depth of this node in the tree (root = 0).

• children List[ThinkNode] - List of child nodes.

• visits int - Number of times this node has been visited during search.

  The node automatically maintains the tree structure by:

  • Setting its depth based on the parent’s depth + 1.
  • Adding itself to the parent’s children list if the parent exists.
  • Providing trajectory utilities to get the full path from root to this node.

​

trajectory

@property
def trajectory() -> str

Get a formatted string representation of the path from root to this node.

Returns:

• str - A formatted string showing the question and each step in the reasoning process

​

backpropagate

def backpropagate(reward: float)

Update the score of this node and its parents using moving average.

​

to_dict

def to_dict() -> dict

Convert ThinkNode to dictionary representation.

Returns:

• Dict - Dictionary containing all node attributes and recursive children

​

from_dict

@classmethod
def from_dict(cls,
              data: dict,
              parent: Optional["ThinkNode"] = None) -> "ThinkNode"

Create ThinkNode from dictionary representation.

Arguments:

• data Dict - Dictionary containing node data
• parent Optional[ThinkNode] - Parent node to attach to

Returns:

• ThinkNode - Reconstructed node with all children

​

visualize_tree

def visualize_tree(root: ThinkNode) -> None

Visualize the tree of thoughts using graphviz.

​

extract_sft_dataset

def extract_sft_dataset(root)

Extract the best trajectory or multiple equally good trajectories for SFT training.

Arguments:

• root - The root node of the tree.

Returns:

List of best trajectories, where each trajectory is a pair of instruction and response.

​

extract_rlhf_preference_dataset

def extract_rlhf_preference_dataset(root, contrastive_threshold=0.2)

Extract and generate preference pairs for RLHF training by comparing sibling nodes.

Arguments:

• root - The root node of the tree.
• contrastive_threshold float - between (0, 1), a distance measure that we are confidence to call one is positive and another is negative.

Returns:

A list of preference pairs, where each pair contains two responses and indicates which one is preferred.

​

ReasoningAgent

class ReasoningAgent(AssistantAgent)

​

__init__

def __init__(name,
             llm_config,
             grader_llm_config=None,
             max_depth=4,
             beam_size=3,
             answer_approach="pool",
             verbose=True,
             reason_config: dict = \{},
             **kwargs) -> None

Initialize a ReasoningAgent that uses tree-of-thought reasoning.

Arguments:

• name - Name of the agent

• llm_config - Configuration for the language model

• grader_llm_config - Optional separate configuration for the grader model. If not provided, uses llm_config

• max_depth int - Maximum depth of the reasoning tree

• beam_size int - DEPRECATED. Number of parallel reasoning paths to maintain

• answer_approach str - DEPRECATED. Either “pool” or “best” - how to generate final answer

• verbose bool - Whether to show intermediate steps

• reason_config dict - Configuration for the reasoning method. Supported parameters:

• method str - The search strategy to use. Options:

  • “beam_search” (default): Uses beam search with parallel paths
  • “mcts”: Uses Monte Carlo Tree Search for exploration
  • “lats”: Uses Language Agent Tree Search with per-step rewards
  • “dfs”: Uses depth-first search (equivalent to beam_search with beam_size=1) Common parameters:

• max_depth int - Maximum depth of reasoning tree (default: 3)

• forest_size int - Number of independent trees to maintain (default: 1)

• rating_scale int - Scale for grading responses, e.g. 1-10 (default: 10)

  Beam Search specific:

• beam_size int - Number of parallel paths to maintain (default: 3)

• answer_approach str - How to select final answer, “pool” or “best” (default: “pool”)

  MCTS/LATS specific:

• nsim int - Number of simulations to run (default: 3)

• exploration_constant float - UCT exploration parameter (default: 1.41)

  Example configs:

• \{"method" - “beam_search”, “beam_size”: 5, “max_depth”: 4}

• \{"method" - “mcts”, “nsim”: 10, “exploration_constant”: 2.0}

• \{"method" - “lats”, “nsim”: 5, “forest_size”: 3}

​

generate_forest_response

def generate_forest_response(messages, sender, config=None)

Generate a response using tree-of-thought reasoning.

Arguments:

• messages - Input messages to respond to
• sender - Agent sending the messages
• config - Optional configuration

Returns:

Tuple[bool, str]: Success flag and generated response

​

rate_node

def rate_node(node: ThinkNode,
              ground_truth: str = None,
              is_outcome: bool = False) -> float

Rate the quality of a reasoning path using the grader agent.

Arguments:

• node ThinkNode - Node containing the reasoning trajectory to evaluate
• is_outcome bool - indicates whether the rating is for an outcome (final answer) or a process (thinking trajectory).

Returns:

• float - Normalized score between 0 and 1 indicating trajectory quality