All experiments were run on Unity. To set up the environment, run:

module load uri/main
module load Java/21.0.2
source /project/pi_miyyer_umass_edu/yixiao/CS646/FinalProject/final_proj_env/bin/activate

Alternatively, you can build your own venv using requirements.txt.

Please refer to this file for details on running baseline experiments and evaluations.

Run zero-shot fixed-n decomposition:

python3 code/query_extraction.py --max_num n

Run few-shot dynamic-n decomposition:

python3 code/query_extraction_fewshot.py

Run a *n-k-m configuration:

python3 code/decomp.py --maxq n --topk k --ndocs m

This code will generate all final contexts and final answers. If you want to use the queries decomposed with the few-shot method, unset --maxq (it defaults to None).

To run incremental retrieval and get final contexts, run:

python3 code/nk.py --steps s --queries n --docs k

To get the final answers with Qwen-2.5-14B-Instruct, run:

python3 code/nk_generate_final_answer.py --steps s --queries n --docs k