mathilda1110/travel-assistant-agent

Collaborative Travel Assistant Agent

A collaborative travel planning assistant built with Smolagents that helps users plan trips through conversation, with consent-based tool usage, feasibility checking, and personalized recommendations.

🎯 Goal

A personalized travel-planning assistant that guides users toward well-informed trip decisions through:

• Collaborative conversation with consent-based tool usage
• Progressive disclosure to avoid overwhelming users
• Feasibility checking before finalizing plans
• User preference modeling for personalized recommendations

🛠️ Tools Available

• user_input_tool – Ask contextual questions and wait for user responses
• infer_task_commonground_from_conversation – Summarize agreed-upon constraints/preferences
• infer_user_profile_from_conversation – Build actionable bullet-point profile of inferred user preferences
• web_search_tool (Bing) – Look up current travel information (max 3 results) with user consent
• visit_webpage_tool – Extract information from specific web pages with user consent
• feasibility_guard (new) – Quick PASS/FAIL on time/budget realism with concrete adjustment suggestions

📋 Key Requirements

ID	Requirement	Linked Principle
R1	Consent before external calls – Ask "Proceed to look up X?"; only search on explicit "yes"	G3, G7
R2	Chunked replies (2–3 bullets) + "Want more?" prompt	G5, G8
R3	Common ground & user model refresh after key info is learned	G1, G3
R4	Feasibility-first – Run feasibility_guard before proposing final plans	G4, G6
R5	Keep tone casual; end each turn with 1 clear next question/choice	G1

🚀 Quick Start

Installation

pip install 'smolagents[litellm]'
pip install 'smolagents[gradio]'
pip install markdownify ddgs

Load the Agent

from smolagents import CodeAgent

# Load from HuggingFace Hub
agent = CodeAgent.from_hub(
    "mathilda1110/travel-assistant-agent",
    trust_remote_code=True
)

# Or load directly from agent.py
from agent import create_travel_assistant
agent = create_travel_assistant()

Use the Agent

# Run a single query
response = agent.run("I want to plan a 4-day trip to Japan with a budget of ¥90,000")

# Or interact in a loop
agent.memory.reset()
while True:
    user_input = input("You: ")
    if user_input.lower() in ['exit', 'quit']:
        break
    response = agent.run(user_input)
    print(f"Agent: {response}")

Using with Gradio UI

from agent import create_travel_assistant
from smolagents.gradio_ui import GradioUI

agent = create_travel_assistant()
GradioUI(agent).launch()

🔧 Configuration

Model Setup

The agent uses AWS Bedrock by default. Set environment variables:

export AWS_ACCESS_KEY_ID=your_key
export AWS_SECRET_ACCESS_KEY=your_secret
export AWS_REGION_NAME=us-east-1
export MODEL_ID=us.meta.llama3-3-70b-instruct-v1:0  # Optional, for custom model

Alternatively, you can use any model supported by LiteLLM by passing a custom model to create_travel_assistant(model=your_model).

Web Search Setup

For web search functionality, you may need to set up Bing API keys:

export BING_SUBSCRIPTION_KEY=your_key

🎨 Key Improvements Over Baseline

1. Consent-Based Tool Usage

• Before: Agent triggered web searches without asking
• After: Agent asks "I can fetch live prices. Proceed? (Yes/No)" before searching
• Benefit: Transparent actions, fewer failed/unused calls

2. Chunked Information Disclosure

• Before: Long, dense replies with many items at once
• After: 2–3 key points at a time, with "Want more details?" prompt
• Benefit: Reduced cognitive load, better user experience

3. Feasibility Guard

• Before: Plans could exceed budget/time without warnings
• After: Automatic feasibility check before finalizing plans with concrete adjustments
• Benefit: Early detection of unrealistic plans, actionable fixes

📚 Framework

Built with Smolagents - a lightweight, code-first agent framework from Hugging Face.

📄 License

Apache 2.0

👤 Author

mathilda1110

---

Note: This agent follows human-centered AI design principles (Microsoft Guidelines for Human-AI Interaction) to provide a better collaborative experience.