ReAct Agent¶
In reality, many complex use cases require combining reasoning and acting advancements to enable language models to solve various reasoning and decision-making tasks. Language models are getting better at both reasoning and acting, but these two directions have largely remained separate.
ReAct enables language models to generate both verbal reasoning traces and textual actions in an interleaved manner. While actions lead to observational feedback from an external environment (referred to as “Env” in the figure below), reasoning traces do not affect the external environment. Instead, they influence the model’s internal state by reasoning over the context and updating it with useful information to support future reasoning and actions.
ReAct is particularly useful in many use cases, such as solving math, logic, coding, and writing problems. That is why Vinagent offers a default ReAct Agent to facilitate their initialization for handling complex tasks that require interleaved reasoning and acting. You can learn more about ReAct in the original paper: ReAct: Synergizing Reasoning and Acting in Language Models, Yao et al., 2022.
Prerequisites¶
Let's install vinagent package and write environment variables to .env file.
Set environment varibles inside .env including TOGETHER_API_KEY to use llm models at togetherai site and TAVILY_API_KEY to use tavily websearch tool at tavily site:
Create an ReAct Agent¶
from langchain_together import ChatTogether
from vinagent.agent.prebuilt import ReactAgent
from dotenv import load_dotenv
load_dotenv()
llm = ChatTogether(
model="meta-llama/Llama-3.3-70B-Instruct-Turbo-Free"
)
To demonstrate the efficient of ReAct Agent, let's assume we need to compute two dog weights average is Husky and Bulldog.
%%writefile vinagent/tools/average_dogs.py
def weight_of_bulldog():
"""The weight of a bulldog"""
return 25
def weight_of_husky():
"""The weight of a husky"""
return 20
def average_weight_of_two_dogs(weight1: float, weight2: float):
"""The average weight of two dogs"""
return (weight1 + weight2) / 2
Let's initialize the ReAct Agent and pass in the tools and the description of the agent.
agent = ReactAgent(
decription="You are a helpful assistant.",
skills=["Search on internet"],
tools=[
'vinagent.tools.average_dogs',
'vinagent.tools.websearch_tools'
],
num_buffered_messages=10,
llm=llm
)
INFO:vinagent.register.tool:Registered weight_of_bulldog:
{'tool_name': 'weight_of_bulldog', 'arguments': {}, 'return': '25', 'docstring': 'The weight of a bulldog', 'dependencies': [], 'module_path': 'vinagent.tools.average_dogs', 'tool_type': 'module', 'tool_call_id': 'tool_154b5559-0e7e-436b-9f9e-741df44e8416'}
INFO:vinagent.register.tool:Registered weight_of_husky:
{'tool_name': 'weight_of_husky', 'arguments': {}, 'return': '20', 'docstring': 'The weight of a husky', 'dependencies': [], 'module_path': 'vinagent.tools.average_dogs', 'tool_type': 'module', 'tool_call_id': 'tool_b4804fad-30cd-4a62-a019-869c70bd86ab'}
INFO:vinagent.register.tool:Registered average_weight_of_two_dogs:
{'tool_name': 'average_weight_of_two_dogs', 'arguments': {'weight1': 0.0, 'weight2': 0.0}, 'return': 'None', 'docstring': 'The average weight of two dogs', 'dependencies': [], 'module_path': 'vinagent.tools.average_dogs', 'tool_type': 'module', 'tool_call_id': 'tool_b77b87ff-1fef-4ee1-8a62-ebbde8dcfc00'}
INFO:vinagent.register.tool:Completed registration for module vinagent.tools.average_dogs
INFO:vinagent.register.tool:Registered search_api:
{'tool_name': 'search_api', 'arguments': {'query': '{}'}, 'return': 'Any', 'docstring': 'Search for an answer from a query string\n Args:\n query (dict[str, str]): The input query to search\n Returns:\n The answer from search query', 'dependencies': ['os', 'dotenv', 'tavily', 'dataclasses', 'typing'], 'module_path': 'vinagent.tools.websearch_tools', 'tool_type': 'module', 'tool_call_id': 'tool_2ec9aaa3-8585-4aec-b5f2-f86335c1f99a'}
INFO:vinagent.register.tool:Completed registration for module vinagent.tools.websearch_tools
Let's ask the average weight of two dogs.
INFO:vinagent.agent.agent:Tool calling iteration 1/10
INFO:vinagent.agent.agent:Executing tool call: {'tool_name': 'weight_of_husky', 'tool_type': 'module', 'arguments': {}, 'module_path': 'vinagent.tools.average_dogs'}
INFO:vinagent.register.tool:Completed executing module tool weight_of_husky({})
INFO:vinagent.agent.agent:Tool calling iteration 2/10
INFO:vinagent.agent.agent:Executing tool call: {'tool_name': 'weight_of_bulldog', 'tool_type': 'module', 'arguments': {}, 'module_path': 'vinagent.tools.average_dogs'}
INFO:vinagent.register.tool:Completed executing module tool weight_of_bulldog({})
INFO:vinagent.agent.agent:Tool calling iteration 3/10
INFO:vinagent.agent.agent:Executing tool call: {'tool_name': 'average_weight_of_two_dogs', 'tool_type': 'module', 'arguments': {'weight1': 20.0, 'weight2': 25.0}, 'module_path': 'vinagent.tools.average_dogs'}
INFO:vinagent.register.tool:Completed executing module tool average_weight_of_two_dogs({'weight1': 20.0, 'weight2': 25.0})
INFO:vinagent.agent.agent:Tool calling iteration 4/10
AIMessage(content='I now know the final answer\nFinal Answer: The average weight of a Husky and a Bulldog is 22.5.', additional_kwargs={'refusal': None}, response_metadata={'token_usage': {'completion_tokens': 28, 'prompt_tokens': 3179, 'total_tokens': 3207, 'completion_tokens_details': None, 'prompt_tokens_details': None, 'cached_tokens': 0}, 'model_name': 'meta-llama/Llama-3.3-70B-Instruct-Turbo-Free', 'system_fingerprint': None, 'finish_reason': 'stop', 'logprobs': None}, id='run--c2183de6-749b-4a3e-9331-74820095dc18-0', usage_metadata={'input_tokens': 3179, 'output_tokens': 28, 'total_tokens': 3207, 'input_token_details': {}, 'output_token_details': {}})
Reasoning over history¶
The ReAct agent can save the user's chat history. This enables the agent to reason over the retained conversational context, which often contains important information in previous interactions. In this case, we set num_buffered_messages=10, meaning it can store up to 10 messages in the chat history. In the following example, we assume the Husky's weight has changed. The ReAct agent will reason over the chat history to determine the new average weight.
answer = agent.invoke(query="If the average weight of a Husky is 30 kilograms, by how many kilograms does the average weight of the two dogs increase?")
answer
INFO:vinagent.agent.agent:Tool calling iteration 1/10
INFO:vinagent.agent.agent:Executing tool call: {'tool_name': 'weight_of_bulldog', 'tool_type': 'module', 'arguments': {}, 'module_path': 'vinagent.tools.average_dogs'}
INFO:vinagent.register.tool:Completed executing module tool weight_of_bulldog({})
INFO:vinagent.agent.agent:Tool calling iteration 2/10
INFO:vinagent.agent.agent:Executing tool call: {'tool_name': 'average_weight_of_two_dogs', 'tool_type': 'module', 'arguments': {'weight1': 30.0, 'weight2': 25.0}, 'module_path': 'vinagent.tools.average_dogs'}
INFO:vinagent.register.tool:Completed executing module tool average_weight_of_two_dogs({'weight1': 30.0, 'weight2': 25.0})
INFO:vinagent.agent.agent:Tool calling iteration 3/10
INFO:vinagent.agent.agent:Executing tool call: {'tool_name': 'average_weight_of_two_dogs', 'tool_type': 'module', 'arguments': {'weight1': 20.0, 'weight2': 25.0}, 'module_path': 'vinagent.tools.average_dogs'}
INFO:vinagent.register.tool:Completed executing module tool average_weight_of_two_dogs({'weight1': 20.0, 'weight2': 25.0})
INFO:vinagent.agent.agent:Tool calling iteration 4/10
INFO:vinagent.agent.agent:Executing tool call: {'tool_name': 'search_api', 'tool_type': 'module', 'arguments': {'query': '27.5 - 22.5'}, 'module_path': 'vinagent.tools.websearch_tools'}
INFO:vinagent.register.tool:Completed executing module tool search_api({'query': '27.5 - 22.5'})
INFO:vinagent.agent.agent:Tool calling iteration 5/10
INFO:vinagent.agent.agent:No more tool calls needed. Completed in 5 iterations.
AIMessage(content='The increase in the average weight of the two dogs is 5 kilograms.\n\nThought: I now know the final answer\nFinal Answer: The average weight of the two dogs increases by 5 kilograms.', additional_kwargs={'refusal': None}, response_metadata={'token_usage': {'completion_tokens': 41, 'prompt_tokens': 3469, 'total_tokens': 3510, 'completion_tokens_details': None, 'prompt_tokens_details': None, 'cached_tokens': 0}, 'model_name': 'meta-llama/Llama-3.3-70B-Instruct-Turbo-Free', 'system_fingerprint': None, 'finish_reason': 'stop', 'logprobs': None}, id='run--4231bdd1-f0c4-492d-a716-af74a144b0ca-0', usage_metadata={'input_tokens': 3469, 'output_tokens': 41, 'total_tokens': 3510, 'input_token_details': {}, 'output_token_details': {}})
Reasoning with search engine¶
Test with another use case that compares the populations of the three cities: New York, Beijing, and Hanoi.
answer = agent.invoke(query="Which city has the greatest population among New York, Beijing, and Hanoi?")
answer
INFO:vinagent.agent.agent:No authentication card provided, skipping authentication
INFO:vinagent.agent.agent:I'am chatting with unknown_user
INFO:vinagent.agent.agent:Tool calling iteration 1/10
INFO:vinagent.agent.agent:Executing tool call: {'tool_name': 'search_api', 'tool_type': 'module', 'arguments': {'query': 'population of New York'}, 'module_path': 'vinagent.tools.websearch_tools'}
INFO:vinagent.register.tool:Completed executing module tool search_api({'query': 'population of New York'})
INFO:vinagent.agent.agent:Tool calling iteration 2/10
INFO:vinagent.agent.agent:Executing tool call: {'tool_name': 'search_api', 'tool_type': 'module', 'arguments': {'query': 'population of Beijing'}, 'module_path': 'vinagent.tools.websearch_tools'}
INFO:vinagent.register.tool:Completed executing module tool search_api({'query': 'population of Beijing'})
INFO:vinagent.agent.agent:Tool calling iteration 3/10
INFO:vinagent.agent.agent:Executing tool call: {'tool_name': 'search_api', 'tool_type': 'module', 'arguments': {'query': 'population of Hanoi'}, 'module_path': 'vinagent.tools.websearch_tools'}
INFO:vinagent.register.tool:Completed executing module tool search_api({'query': 'population of Hanoi'})
INFO:vinagent.agent.agent:Tool calling iteration 4/10
INFO:vinagent.agent.agent:No more tool calls needed. Completed in 4 iterations.
AIMessage(content='We have the populations of New York, Beijing, and Hanoi. New York has a population of approximately 8.5 million, Beijing has a population of around 21.8 million, and Hanoi has a population of around 8,807,523.\n\nThought: Comparing these numbers, we can see that Beijing has the greatest population among the three cities.\n\nThought: I now know the final answer\nFinal Answer: Beijing has the greatest population among the three cities, with a population of approximately 21.8 million.', additional_kwargs={'refusal': None}, response_metadata={'token_usage': {'completion_tokens': 109, 'prompt_tokens': 3331, 'total_tokens': 3440, 'completion_tokens_details': None, 'prompt_tokens_details': None, 'cached_tokens': 0}, 'model_name': 'meta-llama/Llama-3.3-70B-Instruct-Turbo-Free', 'system_fingerprint': None, 'finish_reason': 'stop', 'logprobs': None}, id='run--28a56b66-22bc-420e-8be8-914992b322fe-0', usage_metadata={'input_tokens': 3331, 'output_tokens': 109, 'total_tokens': 3440, 'input_token_details': {}, 'output_token_details': {}})
Test with a math problem that requires reasoning and acting to be interleaved across multiple cycles.
answer = agent.invoke(query="Vinagent is a community project of DataScienceWorld.Kan. It is an initiative by the founder of DataScienceWorld.Kan. Who is the leader of Vinagent?")
answer
INFO:vinagent.agent.agent:No authentication card provided, skipping authentication
INFO:vinagent.agent.agent:I'am chatting with unknown_user
INFO:vinagent.agent.agent:Tool calling iteration 1/10
INFO:vinagent.agent.agent:Executing tool call: {'tool_name': 'search_api', 'tool_type': 'module', 'arguments': {'query': 'DataScienceWorld.Kan founder'}, 'module_path': 'vinagent.tools.websearch_tools'}
INFO:vinagent.register.tool:Completed executing module tool search_api({'query': 'DataScienceWorld.Kan founder'})
INFO:vinagent.agent.agent:Tool calling iteration 2/10
INFO:vinagent.agent.agent:No more tool calls needed. Completed in 2 iterations.
AIMessage(content='Since Pham Đinh Khanh is the founder of DataScienceWorld.Kan and Vinagent is an initiative of this founder, we can conclude that Pham Đinh Khanh is the leader of Vinagent.\n\nThought: I now know the final answer\nFinal Answer: Pham Đinh Khanh is the leader of Vinagent.', additional_kwargs={'refusal': None}, response_metadata={'token_usage': {'completion_tokens': 69, 'prompt_tokens': 2889, 'total_tokens': 2958, 'completion_tokens_details': None, 'prompt_tokens_details': None, 'cached_tokens': 0}, 'model_name': 'meta-llama/Llama-3.3-70B-Instruct-Turbo-Free', 'system_fingerprint': None, 'finish_reason': 'stop', 'logprobs': None}, id='run--b9a36ebc-aa2f-42c5-a6d8-2a3633da41e9-0', usage_metadata={'input_tokens': 2889, 'output_tokens': 69, 'total_tokens': 2958, 'input_token_details': {}, 'output_token_details': {}})