Feedback Models for HyDE

This is the code for the paper: Revisiting Feedback Models for HyDE

Overview

Recent approaches that leverage large language models (LLMs) for pseudo-relevance feedback (PRF) have generally not utilized well-established feedback models like Rocchio and RM3 when expanding queries for sparse retrievers like BM25. Instead, they often opt for a simple string concatenation of the query and LLM-generated expansion content. But is this optimal? To answer this question, we revisit and systematically evaluate traditional feedback models in the context of HyDE, a popular method that enriches query representations with LLM-generated hypothetical answer documents. Our experiments show that HyDE's effectiveness can be substantially improved when leveraging feedback algorithms such as Rocchio to extract and weight expansion terms, providing a simple way to further enhance the accuracy of LLM-based PRF methods.

Setup

We use the following packages:

vllm==0.10.1.1
pyserini==1.2.0
transformers==4.56.1
torch==2.7.1

Run HyDE with feedback models!

To run HyDE with your favorite feedback method, run the following command:

python run_hyde.py \
    --model_path $model \
    --corpus_name $corpus \
    --returned_hits 100 \
    --feedback_mechanism $feedback_mechanism \
    --output_folder $output_folder \

where:

--model_path: the path to your LLM (Code works with Qwen2.5-7B-Instruct, Qwen3-14B, gpt-oss-20b; minor adaptations may be needed for other LLMs)
--corpus_name: Corpus to evaluate on (e.g., 'dl19', 'dl20', 'news', 'covid', 'fiqa', etc; See modules/index_paths.py)
--returned_hits: Number of passages to retrieve with HyDE
--feedback_mechanism: Feedback mechanism for creating HyDE query (Options: 'hyde', 'rocchio', 'rm3', 'mugi', 'naive')
--output_folder: Folder to save HyDE ranking

This general code should extend to additional corpora (as long as its in Pyserini). To do so, just add to THE_SPARSE_INDEX and THE_TOPICS in modules/index_paths.py and add a HyDE prompt to modules/hyde.py.

Running our Query2Doc and generic BM25 pseudo-relevance feedback baselines should be just as simple as the above!

Query2Doc:

python run_query2doc.py \
    --model_path $model \
    --corpus_name $corpus \
    --returned_hits 100 \
    --output_folder $output_folder \

This will run our baseline which generates a single hypothetical passage and repeats the query 5 times.

And, for traditional PRF with BM25:

python run_prf.py \
    --corpus_name $corpus \
    --feedback_documents 8 \
    --returned_hits 100 \
    --feedback_mechanism $feedback_mechanism  \
    --output_folder $output_folder \

If you leave --feedback_mechanism blank, this will run standard BM25. Otherwise, this will run similarly to run_hyde.py

Citation

Please cite our paper if it is helpful to your work!

@article{jedidi2025revisiting,
  title={Revisiting Feedback Models for HyDE},
  author={Jedidi, Nour and Lin, Jimmy},
  journal={arXiv preprint arXiv:2511.19349},
  year={2025}
}