Constrained Auto-Regressive Decoding Constrains Generative Retrieval

Environment setup

We use poetry to manage the dependencies. To install the dependencies, run the following command:

poetry install

Data

We use the semantic ID mapping from Zeng et al. (2024) in their GitHub repository.

Additionally, we use Pyserini to get the run on MSMARCO passage collection. To install Pyserini, follow the instructions in their repository.

Run

To run synthetic experiments, run the following command:

poetry run python synthetic_constraint.py

To run experiments on real data, please following the command within python code named simulation_*.py.