Agent with custom prompt

[ryanshrott]

I'm building my own agent with some custom tools. How can I initially prompt the agent with its main goal and task is? For example, you are a realtor connecting clients with agents etc.

from langchain.agents import initialize_agent, Tool
from langchain.agents import AgentType
tools = [
    Tool(
        name = "model_prices db",
        func=db_chain.run,
        description="useful for when you need to answer questions about model_prices postgres database. Input should be in the form of a question containing full context"
        ),
    Tool(
        name = "current market news",
        func=news.run,
        description="useful for when you need to answer current market news questions about canadian real estate market. Input should be a fully formed question.",
        ),
     Tool(
        name = "wikipedia",
        func=wikipedia.run,
        description="useful for when you need know general information. Input should be a fully formed question.",
     ),
    toolkit.get_tools()[0]
    ] 

from langchain.agents import initialize_agent

conversational_agent = initialize_agent(
    agent='chat-conversational-react-description', 
    tools=tools, 
    llm=llm,
    verbose=True,
    max_iterations=3,
    early_stopping_method="generate",
    memory=memory,
)
     

[jakesteelman]

If you're using a chat LLM, this is similar to what I did.

from langchain.agents import ConversationalChatAgent, AgentExecutor

agent = ConversationalChatAgent.from_llm_and_tools(
     llm=chat_llm,
     tools=tools,
     system_message="this is the prompt/prefix/system message",
     human_message="this is the suffix/human message",
     verbose=True,
     memory=memory,
     # ...
)

agent_chain = AgentExecutor.from_agent_and_tools(
     agent=agent,
     tools=tools,
     memory=memory,
     verbose=True,
     # ...
)

If you look at the reference for the ConversationalChatAgent you can see the default values for syetem_message and human_message and structure your prompt around those. Hope this helps!

[PJBentham]

This worked for me.

What threw me initially is tools, memory & verbose in both agent and agent_chain - do we need to do that?