Agents

In this lesson, you will learn how to create an agent.

Agents wrap a model and give it access to a set of tools. These tools may access additional data sources, APIs, or functionality. The model is used to determine which of the tools to use to complete a task.

The agent you will create will be able to chat about movies and search YouTube for movie trailers.

Tools

A tool is a specific abstraction around a function that makes it easy for a language model to interact with it. Langchain provides several tools out of the box, and you can create tools to extend the functionality of your agents.

You will use the YouTube Tool to search YouTube for movie trailers.

Movie trailer agent

Review the program below, before running it.

python
from langchain_openai import ChatOpenAI
from langchain.agents import AgentExecutor, create_react_agent
from langchain.tools import Tool
from langchain import hub
from langchain_core.prompts import ChatPromptTemplate
from langchain_core.runnables.history import RunnableWithMessageHistory
from langchain.schema import StrOutputParser
from langchain_community.chat_message_histories import Neo4jChatMessageHistory
from langchain_community.graphs import Neo4jGraph
from uuid import uuid4

SESSION_ID = str(uuid4())
print(f"Session ID: {SESSION_ID}")

llm = ChatOpenAI(openai_api_key="sk-...")

graph = Neo4jGraph(
    url="bolt://localhost:7687",
    username="neo4j",
    password="pleaseletmein"
)

prompt = ChatPromptTemplate.from_messages(
    [
        (
            "system",
            "You are a movie expert. You find movies from a genre or plot.",
        ),
        ("human", "{input}"),
    ]
)

movie_chat = prompt | llm | StrOutputParser()

def get_memory(session_id):
    return Neo4jChatMessageHistory(session_id=session_id, graph=graph)

tools = [
    Tool.from_function(
        name="Movie Chat",
        description="For when you need to chat about movies. The question will be a string. Return a string.",
        func=movie_chat.invoke,
    )
]

agent_prompt = hub.pull("hwchase17/react-chat")
agent = create_react_agent(llm, tools, agent_prompt)
agent_executor = AgentExecutor(agent=agent, tools=tools)

chat_agent = RunnableWithMessageHistory(
    agent_executor,
    get_memory,
    input_messages_key="input",
    history_messages_key="chat_history",
)

while True:
    q = input("> ")

    response = chat_agent.invoke(
        {
            "input": q
        },
        {"configurable": {"session_id": SESSION_ID}},
    )
    
    print(response["output"])
Click here to see a typical output from this program
[user] Find a movie where aliens land on earth.
[chat model] Sure, I can help you with that. One movie I would recommend where aliens land on Earth is "Arrival" (2016). It's a science fiction film directed by Denis Villeneuve. The story follows a linguist who is recruited by the military to help communicate with an alien species that has landed on Earth. It's a thought-provoking and visually stunning movie that explores themes of communication, time, and the human experience. I hope you enjoy it!

Based on your understanding from previous lessons, you should be able to identify the following:

  1. A chat model is being used to have a conversation about movies

  2. The prompt which sets the context for the LLM and the input variables

  3. That memory is used to store the conversation history in a Neo4j database

In addition to the above, the following is new:

  1. A tool is created using the chain:

    python
    tools = [
        Tool.from_function(
            name="Movie Chat",
            description="For when you need to chat about movies. The question will be a string. Return a string.",
            func=movie_chat.invoke,
        )
    ]
  2. An agent is created that uses the tool:

    python
    agent_prompt = hub.pull("hwchase17/react-chat")
    agent = create_react_agent(llm, tools, agent_prompt)
    agent_executor = AgentExecutor(agent=agent, tools=tools)
  3. The agent is wrapped in a RunnableWithMessageHistory chain that allows it to interact with the memory:

    python
    chat_agent = RunnableWithMessageHistory(
        agent_executor,
        get_memory,
        input_messages_key="input",
        history_messages_key="chat_history",
    )

Add your openai_api_key and update the Neo4jGraph connection details to run the program.

Click to reveal your Sandbox connection details
Connection URL

bolt://{sandbox_ip}:{sandbox_boltPort}

Username

{sandbox_username}

Password

{sandbox_password}

Creating tools

Tools are interfaces that an agent can interact with. You can create custom tools able to perform any functionality you want.

In this example, the Tool is created from a function. The function is the movie_chat.invoke method.

python
tools = [
    Tool.from_function(
        name="Movie Chat",
        description="For when you need to chat about movies. The question will be a string. Return a string.",
        func=movie_chat.invoke,
    )
]

The name and description help the LLM select which tool to use when presented with a question. The func parameter is the function that will be called when the tool is selected. The return_direct flag indicates that the tool will return the result directly.

Agents support multiple tools, so you pass them to the agent as a list (tools).

Initializing an agent

The following code creates the agent:

python
agent_prompt = hub.pull("hwchase17/react-chat")
agent = create_react_agent(llm, tools, agent_prompt)
agent_executor = AgentExecutor(agent=agent, tools=tools)

There are different types of agents that you can create. This example creates a ReAct - Reasoning and Acting) agent type.

An agent requires a prompt. You could create a prompt, but in this example, the program pulls a pre-existing prompt from the Langsmith Hub.

The hwcase17/react-chat prompt instructs the model to provide an answer using the tools available in a specific format.

The create_react_agent function creates the agent and expects the following parameters:

  • The llm that will manage the interactions and decide which tool to use

  • The tools that the agent can use

  • The prompt that the agent will use

The AgentExecutor class runs the agent. It expects the following parameters:

  • The agent to run

  • The tools that the agent can use

  • The memory which will store the conversation history

AgentExecutor parameters

You may find the following additional parameters useful when initializing an agent:

  • max_iterations - the maximum number of iterations to run the LLM for. This is useful in preventing the LLM from running for too long or entering an infinite loop.

  • verbose - if True the agent will print out the LLM output and the tool output.

  • handle_parsing_errors - if True the agent will handle parsing errors and return a message to the user.

python
agent_executor = AgentExecutor(
    agent=agent,
    tools=tools,
    max_iterations=3,
    verbose=True,
    handle_parse_errors=True
)

Multiple tools

A key advantage of using an agent is that they can use multiple tools. Access to multiple tools allows you to create agents that can perform several specific tasks.

You can extend this example to allow it to search YouTube for movie trailers by adding the YouTube Tool to the tools list.

Firstly, you will need to install the youtube-search package.

bash
pip install youtube-search

Import the YouTubeSearchTool and create a new tool.

python
from langchain_community.tools import YouTubeSearchTool

youtube = YouTubeSearchTool()

The YouTubeSearchTool tool expects a search term and the number of results passed as a comma-separated string.

The agent may pass queries containing commas, so create a function to strip the commas from the query and pass the query to the YouTubeSearchTool.

python
def call_trailer_search(input):
    input = input.replace(",", " ")
    return youtube.run(input)

Finally, add the call_trailer_search function to the tools list.

python
tools = [
    Tool.from_function(
        name="Movie Chat",
        description="For when you need to chat about movies. The question will be a string. Return a string.",
        func=movie_chat.invoke,
    ),
    Tool.from_function(
        name="Movie Trailer Search",
        description="Use when needing to find a movie trailer. The question will include the word trailer. Return a link to a YouTube video.",
        func=call_trailer_search,
    ),
]
Click here to reveal the complete program
python
from langchain_openai import ChatOpenAI
from langchain.agents import AgentExecutor, create_react_agent
from langchain.tools import Tool
from langchain import hub
from langchain_core.prompts import ChatPromptTemplate
from langchain_core.runnables.history import RunnableWithMessageHistory
from langchain.schema import StrOutputParser
from langchain_community.tools import YouTubeSearchTool
from langchain_community.chat_message_histories import Neo4jChatMessageHistory
from langchain_community.graphs import Neo4jGraph
from uuid import uuid4

SESSION_ID = str(uuid4())
print(f"Session ID: {SESSION_ID}")

llm = ChatOpenAI(openai_api_key="sk-...")

graph = Neo4jGraph(
    url="bolt://localhost:7687",
    username="neo4j",
    password="pleaseletmein"
)

prompt = ChatPromptTemplate.from_messages(
    [
        (
            "system",
            "You are a movie expert. You find movies from a genre or plot.",
        ),
        ("human", "{input}"),
    ]
)

movie_chat = prompt | llm | StrOutputParser()

youtube = YouTubeSearchTool()

def get_memory(session_id):
    return Neo4jChatMessageHistory(session_id=session_id, graph=graph)

def call_trailer_search(input):
    input = input.replace(",", " ")
    return youtube.run(input)

tools = [
    Tool.from_function(
        name="Movie Chat",
        description="For when you need to chat about movies. The question will be a string. Return a string.",
        func=movie_chat.invoke,
    ),
    Tool.from_function(
        name="Movie Trailer Search",
        description="Use when needing to find a movie trailer. The question will include the word trailer. Return a link to a YouTube video.",
        func=call_trailer_search,
    ),
]

agent_prompt = hub.pull("hwchase17/react-chat")
agent = create_react_agent(llm, tools, agent_prompt)
agent_executor = AgentExecutor(agent=agent, tools=tools)

chat_agent = RunnableWithMessageHistory(
    agent_executor,
    get_memory,
    input_messages_key="input",
    history_messages_key="chat_history",
)

while True:
    q = input("> ")

    response = chat_agent.invoke(
        {
            "input": q
        },
        {"configurable": {"session_id": SESSION_ID}},
    )
    
    print(response["output"])

The model will use the name and description for each tool to decide which tool to use.

When prompted to find a movie trailer, the model should use the YouTubeSearchTool tool.

[user] Find the movie trailer for the Matrix.
[agent] Here are the movie trailers for "The Matrix":
The Matrix - Official Trailer #1 -
    https://www.youtube.com/watch?v=vKQi3bBA1y8&pp=ygUKVGhlIE1hdHJpeA%3D%3D
The Matrix - Official Trailer #2 -
    https://www.youtube.com/watch?v=xrYg_qKX-aI&pp=ygUKVGhlIE1hdHJpeA%3D%3D

However, when asked about movies, genres or plots, the model will use the chat_chain tool.

[user] Find a movie about the meaning of life
[agent] Certainly! One movie that explores the meaning of life is "The Tree of Life" directed by Terrence Malick. It follows the journey of a young boy as he grows up in the 1950s and reflects on his experiences and the meaning of existence. It's a visually stunning and thought-provoking film that delves into existential questions.

As the agent also uses the conversation memory, you can refer back to the previous questions, such as finding a trailer for a movie it has recommended:

[user] Can you find the trailer
[agent] Here are two links to the trailer for "The Tree of Life":
Link 1 - https://www.youtube.com/watch?v=RrAz1YLh8nY&pp=ygUQVGhlIFRyZWUgb2YgTGlmZQ%3D%3D
Link 2 - https://www.youtube.com/watch?v=cv-dH5gHi1c&pp=ygUQVGhlIFRyZWUgb2YgTGlmZQ%3D%3D

Agents and tools allow you to create more adaptable and flexible models to perform multiple tasks.

Check Your Understanding

Agents

Select all relating to agents that are True.

  • ✓ Agents give a model access to a set of tools

  • ✓ The model decides which tool to use

  • ❏ Memory is not accessible by an agent

  • ❏ Agents can only use tools provided by Langchain

Hint

Agents are highly flexible and can use a variety of tools to accomplish their goals.

Solution

The correct answers are:

  • Agents give a model access to a set of tools

  • The model decides which tool to use

The others are false because:

  • Agents can have memory

  • You can create custom tools for agents to use

Summary

In this lesson, you learned how to create an agent to use multiple tools.

In the next lesson, you will learn how to use Neo4j as a vector store using Langchain Receivers.