AI Travel Agent

AI travel agent built with LangGraph + LangChain, with Chroma memory (persistent + in-memory), plus step-aware logging and run metrics.

Requirements

• Python 3.11+
• Ollama running locally (OLLAMA_BASE_URL default: http://localhost:11434) or a free Groq API key
• A chat model pulled in Ollama (default: qwen2.5:7b-instruct) or a Groq model (default: llama-3.1-8b-instant)
• (Optional) Amadeus free tier keys for Top‑5 live flights/hotels

Setup

python -m venv .venv
source .venv/bin/activate
pip install -e ".[dev,embeddings]"
cp .env.example .env

Run

ai-travel-agent

Documentation

See:

• docs/SETUP.md for installation + environment setup
• docs/PROMPTS.md for example prompts
• docs/ARCHITECTURE.md for a full architecture walkthrough (graph flow, memory, tools, logging, metrics, evaluation)

Options:

• --log-level INFO|DEBUG
• --runtime-dir ./runtime
• --verbose

LLM Providers

By default the app uses local Ollama. To use a free API instead:

# .env
LLM_PROVIDER=groq
GROQ_API_KEY=your_key_here
GROQ_MODEL=llama-3.1-8b-instant

To switch back to local:

LLM_PROVIDER=ollama

Live Top‑5 Flights/Hotels (Optional)

If you want actual Top‑5 results (not just provider links), add Amadeus keys:

AMADEUS_CLIENT_ID=your_id
AMADEUS_CLIENT_SECRET=your_secret
AMADEUS_BASE_URL=https://test.api.amadeus.com

Artifacts:

• Logs: runtime/logs/app.jsonl and runtime/logs/app.log
• Metrics: runtime/metrics/metrics.jsonl
• Calendar export: runtime/artifacts/*.ics

Notes

• This MVP does not book flights/hotels. It produces a plan, deep links, and an .ics itinerary export.
• Weather and other network tools degrade gracefully if offline.

Evaluation

Each run is evaluated with:

• Hard gates (must pass): assumptions coverage, no fabricated prices, valid links, valid ICS, safety disclaimer
• Rubric scores (0–5): relevance, feasibility, completeness, specificity, coherence

The final evaluation is included in runtime/metrics/metrics.jsonl for each run.