MultiAgentScript

# This script demonstrates how to use a multi-LLM agent framework to perform complex tasks that involve external tools
# The framework consists of three components: a planner, a caller, and a summarizer
# Each component is implemented by a single LLM that is fine-tuned on a specific sub-task
# The framework is based on the approach outlined in the paper "Small LLMs Are Weak Tool Learners: A Multi-LLM Agent" [^1^][1]

# Import the necessary libraries
import requests
import json
import numpy as np

# Define the task and the tools
task = "Find the best restaurant near me based on the ratings and reviews"
tools = ["Google Maps API", "Yelp API", "Sentiment Analysis API"]

# Initialize the planner, the caller, and the summarizer
# The planner is responsible for generating a plan of action, such as which tools to use and in what order
# The caller is responsible for invoking the tools and parsing the results
# The summarizer is responsible for generating a concise and informative summary of the final outcome
planner = LLM("planner")
caller = LLM("caller")
summarizer = LLM("summarizer")

# Fine-tune the LLMs on the respective sub-tasks using a two-stage training paradigm
# First, fine-tune a backbone LLM on the entire dataset without discriminating sub-tasks
# Second, use the fine-tuned LLM to instantiate the planner, caller, and summarizer respectively
# Continually fine-tune them on respective sub-tasks
backbone = LLM("backbone")
backbone.fine_tune(dataset)
planner.copy_from(backbone)
caller.copy_from(backbone)
summarizer.copy_from(backbone)
planner.continue_fine_tune(planner_dataset)
caller.continue_fine_tune(caller_dataset)
summarizer.continue_fine_tune(summarizer_dataset)

# Generate a plan of action using the planner
plan = planner.generate_plan(task, tools)
print(f"Plan: {plan}")

# Execute the plan using the caller
results = []
for tool in plan:
    result = caller.invoke_tool(tool, task)
    results.append(result)
    print(f"Result from {tool}: {result}")

# Generate a summary using the summarizer
summary = summarizer.generate_summary(results, task)
print(f"Summary: {summary}")