Skip to main content

Structured Tool Chat Agent

Overview​

The structured tool chat agent can utilize multiple tools using their args_schema to populate the action input. This allows more complex multi-tool interactions compared to agents that take a single string action input.

Configuration​

Import required modules:

from langchain.agents import AgentType  
from langchain.chat_models import ChatOpenAI
from langchain.agents import initialize_agent

Initialize tools - here we use a browser, Google Sheets, and a SQL database:

browser_toolkit = PlaywrightBrowserToolkit(...)
spreadsheet_toolkit = GoogleSheetsToolkit(...)
sql_toolkit = PostgresToolkit(...)
tools = browser_toolkit.get_tools() + spreadsheet_toolkit.get_tools() + sql_toolkit.get_tools()

Initialize the agent with the tools:

agent = initialize_agent(tools, llm, agent_type=AgentType.STRUCTURED_TOOL_CHAT)

Usage​

Call the agent to take multi-tool actions:

response = await agent.run("Browse to example.com, add the page title to my spreadsheet, and insert the URL into my database")

The agent will use the tools' args_schema to populate the action input.

Examples​

Web Scraping and Data Storage​

Scrape a website and store data in a database:

input = "Browse to example.com, extract the page text, and store it in my database"
response = await agent.run(input)

The agent will:

  1. Navigate to the website using the browser tool
  2. Extract the page text using the browser tool
  3. Insert the text into the configured database using the SQL tool

Web Automation and Spreadsheets​

Automate form submissions and track data in a spreadsheet:

input = "Go to example.com/form, fill in the form fields with my details, submit it, and add the submission timestamp to my spreadsheet" 
response = await agent.run(input)

The agent will:

  1. Navigate to the form using the browser tool
  2. Populate the form fields using the browser tool
  3. Submit the form using the browser tool
  4. Add the submission timestamp to the configured spreadsheet using the Google Sheets tool

Complex Orchestration​

Chain together actions across all tools:

input = "Browse to example.com, extract the links, store them in my database, add the link count to my spreadsheet, and summarize the page content"
response = await agent.run(input)

How args_schema populates the action input​

The args_schema defined by each tool specifies the required parameters. For example:

# Browser tool args_schema
browser_args_schema = {"url": "The URL to navigate to"} 

# Spreadsheet tool args_schema
spreadsheet_args_schema = {"text": "The text to add", "cell": "The cell reference"}

# SQL tool args_schema 
sql_args_schema = {"text": "The text to insert", "table": "The table name"}

The agent uses these schemas to populate the action input:

# Action input for browser tool
action_input = {"action": "navigate_browser", 
                "action_input": {"url": "https://example.com"}}

# Action input for spreadsheet tool                 
action_input = {"action": "add_text_cell",
                "action_input": {"text": "Timestamp", "cell": "A1"}}

# Action input for SQL tool
action_input = {"action": "insert_text_table",
                "action_input": {"text": "Page content", "table": "pages"}}

Conclusion​

The structured tool chat agent enables complex orchestration and automation by using multiple tools and their args_schemas. This is an improvement over single action string agents. The agent makes it easy to chain together multi-step workflows across diverse tools.