Skip to content

README.md

Latest commit

 

History

History
77 lines (50 loc) · 2.9 KB

README.md

File metadata and controls

77 lines (50 loc) · 2.9 KB

Find and book plane tickets using LFM2.5-1.2B-Thinking

Discord

This is a minimal Python CLI that helps you find and book plane tickets using tool calling and reasoning to solve multi-step workflows.

For example: Book the cheapest flight from Barcelona to Belgrade on 2026-01-31

This project showcases the power of LFM2.5-1.2B-Thinking, a small Language Model that excells at tasks that require reasoning, logic and strong tool calling skills. And the best part? The model can run on edge devices.

See it in action

Table of contents

Quickstart

  1. Make sure you have uv installed in your system

    uv --version

    If the previous command fails, install uv following these instructions.

  2. Build the project

    uv sync

  3. Ask the flight search assistant to help you find and book plane tickets. For example:

    # single tool call to `search_flights`
uv run flight_search.py --query "What flights are available from New York to Paris on 2026-01-19?"

# single tool call to `book_flight`
uv run flight_search.py --query "Book flight AA495 for 2026-02-04"

# 2-step sequential tool call to `search_flights` and then `book_flight`
uv run flight_search.py --query "Book the cheapest flight from Barcelona to Belgrade on 2026-01-31"

# N-step sequential tool call
uv run flight_search.py --query "Book the cheapest flight from Barcelona to a US city on the East Coast that is not NYC on 2026-02-14"

How does it work?

The model has access to 2 tools:

  • search_flights -> to retrieve contextual information
  • book_flights -> to act on the outside world

Note: The search_flights and book_flight functions are mocked using synthetic data for demonstration purposes. You can integrate with real flight data APIs (e.g., Amadeus, Skyscanner, or Kiwi) for production use.

Given a user request, for example Book the cheapest flight from Barcelona to Belgrade on 2026-01-31 the model iteratively

  • Generate a response possibly with tool calls
  • Executes any tool calls
  • Regenerates response

The model is served fully locally using llama.cpp using the

Next steps

  • Add an evaluation dataset and loop.
  • If necessary, boost model performance with GRPO fine-tuning with verifiable rewards.

Need help?

Join the Liquid AI Discord Community and ask.

Discord