Practical overview
When it fits
Use Travelplanner when the goal is to measure or compare an agent’s planning behavior, especially where a travel query combines environment, commonsense and hard constraints. Choose the two-stage mode for tool-use evaluation or sole-planning mode when isolating planning ability.
Who it is for
- AI researchers studying language-agent planning
- Developers comparing models on tool-using travel tasks
- Teams creating regression benchmarks for planning-agent updates
- Researchers analyzing failures in itinerary generation
Documented uses
- Evaluate an agent’s ability to plan transportation, meals, attractions and accommodation across multiple days.
- Compare supported language models on tool-using and sole-planning travel tasks.
- Convert generated natural-language plans into structured files for evaluation.
- Run offline validation checks on agent-produced travel plans.
Requirements & limits
The repository is public and lists an MIT license. The documented workflows require a local Python environment, the project database and provider credentials for the selected model.
- Two-stage mode evaluates information gathering with search tools before planning.
- Sole-planning mode focuses on planning without the same information-gathering stage.
- The post-processing workflow uses GPT-4 to convert natural-language plans into JSON formats.
- Offline validation is supported through the provided evaluation script, while test-set evaluation is directed to the official leaderboard.
- The repository provides a format-check tool for test-set submission files.
Check current requirements in the official documentation. Source-based guidance does not mean OpenViST has run or independently tested this project.
Set up and evaluate a Travelplanner validation run
Prepare the repository, generate a travel plan, convert it to the required JSON format and evaluate it on the offline validation set.
Before you start
- A public copy of the Travelplanner repository
- Conda with Python 3.9
- The project dependencies from requirements.txt
- The downloaded project database placed in the TravelPlanner directory
- A supported model and the corresponding provider API key
Steps
Create and activate a Conda environment with Python 3.9, then install the repository dependencies.
Download and unzip the database into the TravelPlanner directory.
Run either the two-stage tool-agent workflow or the sole-planning workflow for the validation set, specifying a model and output directory.
Run the post-processing scripts to parse plans, extract elements and combine the resulting files for evaluation.
Run the provided evaluation script against the resulting validation evaluation file.
Expected outcome
A generated travel plan is transformed into the project’s evaluation format and assessed on the offline validation set.
Guide limitations
- The documented workflow requires provider credentials for the selected model.
- The repository directs test-set evaluation to its official leaderboard to avoid data contamination.
- The documented setup does not establish that a particular model or run will produce a successful result.
Sources & review
Source record details
- Source availability
- Active
- Sources checked
- 4 October 2026
- Original website title
- OSU-NLP-Group/TravelPlanner · Official homepage title · retained in its source language