Revisiting Text Ranking in Deep Research

This repository contains the code for the paper Revisiting Text Ranking in Deep Research. This work has been accepted at SIGIR 2026, the 49th International ACM SIGIR Conference on Research and Development in Information Retrieval.

In this work, we reproduce a comprehensive set of text ranking methods in the context of deep research. Specifically, we investigate the performance of 2 deep research agents accessing 5 retrievers and 3 re-rankers. The experiments use the agents gpt-oss-20b and GLM-4.7-Flash (30B). Retrievers are BM25, SPLADE-v3, RepLLaMA, Qwen3-Embedding-8B, and ColBERTv2. Re-rankers are monoT5-3B, RankLLaMA-7B, and Rank1-7B.

If you find this work useful, please consider citing:

@inproceedings{meng2026revisiting,
  title={Revisiting Text Ranking in Deep Research},
  author={Meng, Chuan and Ou, Litu and MacAvaney, Sean and Dalton, Jeff},
  booktitle={Proceedings of the 49th International ACM SIGIR Conference on Research and Development in Information Retrieval},
  pages = {3006--3016},
  url = {https://doi.org/10.1145/3805712.3808557},
  doi = {10.1145/3805712.3808557},
  year={2026}
}

Resources released

To support reproducibility and further analysis, this repository releases the following on the Hugging Face Hub:

The execution trajectories are released in encrypted form and can be decrypted locally using the provided scripts.

Repository Structure

This repository is organised into the following sections:

• 1. Prerequisites
• 2. Quick start
• 3. Reproducing results
  • 3.1 RQ1: Retrievers on passage and document corpora
  • 3.2 RQ2: Re-rankers
  • 3.3 RQ3: Query-to-question (Q2Q) reformulation
• 4. Passage corpus construction
• 5. Indexing
  • 5.1 Passage corpus
  • 5.2 Document corpus
• 6. Decrypting encrypted runs
• 7. Contact

1. Prerequisites

1.1 Environment

We use different environments for (i) hosting local vLLM servers and (ii) running the deep-research agents.

For hosting local vLLM servers for gpt-oss-20b and GLM-4.7-Flash (30B):

uv venv vllm_server --python 3.12
source vllm_server/bin/activate

uv pip install "vllm==0.15.0"
uv pip install -U transformers
uv pip install "numpy==2.2.2" "numba==0.61.2" "llvmlite==0.44.0"

For running the agents, we follow the environment provided in the BrowseComp-Plus repository. Please follow the installation instructions in that repository. After installation, activate the environment using:

source .venv/bin/activate

We use PyLate to run ColBERTv2 , which conflicts with the default BrowseComp-Plus environment. Therefore, we create a separate environment for PyLate:

uv venv pylate --python 3.10
source pylate/bin/activate

uv pip install pylate
uv pip install pytrec_eval
uv pip install pyserini
uv pip install peft
uv pip install qwen_omni_utils
uv pip install rich

1.2 Data downloading

Please follow the instructions in the BrowseComp-Plus repository to download the dataset and generate the decrypted data. This process will produce the decrypted dataset at ./data/browsecomp_plus_decrypted.jsonl, the query file at ./topics-qrels/queries.tsv, and two relevance-judgment files at ./topics-qrels/qrel_evidence.txt and ./topics-qrels/qrel_golds.txt.

Rename the query file as follows:

mv ./topics-qrels/queries.tsv ./topics-qrels/queries-all.tsv

Download the passage corpus

Download the BrowseComp-Plus passage corpus from here:

bash ./scripts_build_index/download_passage.sh

Passage files will be put in the ./data/browsecomp-plus-passage/ directory under the repository root. If you prefer to construct the passage corpus yourself, follow Section 4.

Download retrieval indices

Download all pre-built retriever indices from here:

bash ./scripts_build_index/download_indexes.sh

All indexes will be put in the indexes/ directory under the repository root. If you prefer to build the indices yourself, follow Section 5.

2. Quick start

We begin by replicating the best-performing configurations reported in our paper: the gpt-oss-20b agent using a pipeline with BM25 as the retriever and monoT5-3B as the re-ranker (with a re-ranking depth of 50) on our passage corpus, which achieves a Recall of 0.716 and an answer accuracy of 0.689.

First, launch a local gpt-oss-20b server using vLLM:

source vllm_server/bin/activate

CUDA_DEVICE_ORDER=PCI_BUS_ID CUDA_VISIBLE_DEVICES=0 \
vllm serve openai/gpt-oss-20b \
--tool-call-parser openai --enable-auto-tool-choice \
--host 0.0.0.0 --port 8000

Then, run the agent with BM25 + monoT5:

source .venv/bin/activate

CUDA_DEVICE_ORDER=PCI_BUS_ID CUDA_VISIBLE_DEVICES=1 \
python search_agent/oss_client.py \
  --model openai/gpt-oss-20b \
  --model-url http://localhost:8000/v1 \
  --searcher-type bm25 \
  --index-path ./indexes/index.bm25.passage/ \
  --output-dir ./runs/gpt-oss-20b-high/queries-all.bm25-d50-monot5-3b-msmarco-k5.passage-psgid \
  --reasoning-effort high --max-tokens 40000 --query-template QUERY_TEMPLATE_NO_GET_DOCUMENT \
  --snippet-max-tokens 512 \
  --query ./topics-qrels/queries-all.tsv \
  --reranking-depth 50 --k 5 \
  --reranker-type monot5 \
  --monot5-model castorini/monot5-3b-msmarco --monot5-tokenizer castorini/monot5-3b-msmarco \
  --monot5-batch-size 8 \
  --num-threads 10 

# the Max-P strategy, which maps retrieved passages to documents by assigning each document the maximum score among its retrieved passages
python psg2doc.py \
  --input_json_dir ./runs/gpt-oss-20b-high/queries-all.bm25-d50-monot5-3b-msmarco-k5.passage-psgid \
  --output_json_dir ./runs/gpt-oss-20b-high/queries-all.bm25-d50-monot5-3b-msmarco-k5.passage

Last, run the following commands to do evaluation:

# Search calls, Recall, and answer accuracy
CUDA_DEVICE_ORDER=PCI_BUS_ID CUDA_VISIBLE_DEVICES=1,2 \
python scripts_evaluation/evaluate_run.py \
--input_dir ./runs/gpt-oss-20b-high/queries-all.bm25-d50-monot5-3b-msmarco-k5.passage \
--ground_truth ./data/browsecomp_plus_decrypted.jsonl \
--qrel_evidence ./topics-qrels/qrel_evidence.txt \
--tensor_parallel_size 2

# completion rate
python scripts_evaluation/count_complete.py \
--input_dir ./runs/gpt-oss-20b-high/queries-all.bm25-d50-monot5-3b-msmarco-k5.passage

3. Reproducing results

Before running the experiments, ensure that the required retrieval indices are available.
You may either download our pre-built indices or build them by following the instructions in Section 5.

Next, launch an agent server required for the experiments. Launch gpt-oss-20b:

source vllm_server/bin/activate
bash ./run_scripts/server_gpt-oss-20b.sh

Launch GLM-4.7-Flash:

source vllm_server/bin/activate
bash ./run_scripts/server_GLM-4.7-Flash.sh

You may change the port value in the server scripts if needed. Make sure that the MODEL_URL specified in each experiment script (e.g., MODEL_URL=http://localhost:8000/v1) points to the correct server endpoint you intend to use.

3.1 RQ1: Retrievers on passage and document corpora

All scripts assume that the corresponding indices have been built and the agent server is already running.

gpt-oss-20b

Run the gpt-oss-20b agent with BM25 on
(i) the passage corpus,
(ii) the document corpus, and
(iii) the document corpus with the full-document reader enabled:

source .venv/bin/activate
bash ./run_scripts/rq1/gpt-oss-20b.queries-all.bm25-k5.passage.sh
bash ./run_scripts/rq1/gpt-oss-20b.queries-all.bm25-k5.document.sh
bash ./run_scripts/rq1/gpt-oss-20b.queries-all.bm25-k5.document.getdoc.sh

Run the gpt-oss-20b agent with SPLADE-v3 on the three settings:

source .venv/bin/activate
bash ./run_scripts/rq1/gpt-oss-20b.queries-all.splade-v3-k5.passage.sh
bash ./run_scripts/rq1/gpt-oss-20b.queries-all.splade-v3-k5.document.sh
bash ./run_scripts/rq1/gpt-oss-20b.queries-all.splade-v3-k5.document.getdoc.sh

Run the gpt-oss-20b agent with RepLLaMA on the three settings:

source .venv/bin/activate
bash ./run_scripts/rq1/gpt-oss-20b.queries-all.repllama-v1-7b-lora-passage-k5.passage.sh
bash ./run_scripts/rq1/gpt-oss-20b.queries-all.repllama-v1-7b-lora-passage-k5.document.sh
bash ./run_scripts/rq1/gpt-oss-20b.queries-all.repllama-v1-7b-lora-passage-k5.document.getdoc.sh

Run the gpt-oss-20b agent with Qwen3-Embedding-8B on the three settings:

source .venv/bin/activate
bash ./run_scripts/rq1/gpt-oss-20b.queries-all.qwen3-embedding-8b-k5.passage.sh
bash ./run_scripts/rq1/gpt-oss-20b.queries-all.qwen3-embedding-8b-k5.document.sh
bash ./run_scripts/rq1/gpt-oss-20b.queries-all.qwen3-embedding-8b-k5.document.getdoc.sh

Run the gpt-oss-20b agent with ColBERTv2 on the three settings:

source pylate/bin/activate
bash ./run_scripts/rq1/gpt-oss-20b.queries-all.colbertv2.0-k5.passage.sh
bash ./run_scripts/rq1/gpt-oss-20b.queries-all.colbertv2.0-k5.document.sh
bash ./run_scripts/rq1/gpt-oss-20b.queries-all.colbertv2.0-k5.document.getdoc.sh

GLM-4.7-Flash

Run the GLM-4.7-Flash agent with BM25 on
(i) the passage corpus,
(ii) the document corpus, and
(iii) the document corpus with the full-document reader enabled:

source .venv/bin/activate
bash ./run_scripts/rq1/GLM-4.7-Flash.queries-all.bm25-k5.passage.sh
bash ./run_scripts/rq1/GLM-4.7-Flash.queries-all.bm25-k5.document.sh
bash ./run_scripts/rq1/GLM-4.7-Flash.queries-all.bm25-k5.document.getdoc.sh

Run the GLM-4.7-Flash agent with SPLADE-v3 on the three settings:

source .venv/bin/activate
bash ./run_scripts/rq1/GLM-4.7-Flash.queries-all.splade-v3-k5.passage.sh
bash ./run_scripts/rq1/GLM-4.7-Flash.queries-all.splade-v3-k5.document.sh
bash ./run_scripts/rq1/GLM-4.7-Flash.queries-all.splade-v3-k5.document.getdoc.sh

Run the GLM-4.7-Flash agent with RepLLaMA on the three settings:

source .venv/bin/activate
bash ./run_scripts/rq1/GLM-4.7-Flash.queries-all.repllama-v1-7b-lora-passage-k5.passage.sh
bash ./run_scripts/rq1/GLM-4.7-Flash.queries-all.repllama-v1-7b-lora-passage-k5.document.sh
bash ./run_scripts/rq1/GLM-4.7-Flash.queries-all.repllama-v1-7b-lora-passage-k5.document.getdoc.sh

Run the GLM-4.7-Flash agent with Qwen3-Embedding-8B on the three settings:

source .venv/bin/activate
bash ./run_scripts/rq1/GLM-4.7-Flash.queries-all.qwen3-embedding-8b-k5.passage.sh
bash ./run_scripts/rq1/GLM-4.7-Flash.queries-all.qwen3-embedding-8b-k5.document.sh
bash ./run_scripts/rq1/GLM-4.7-Flash.queries-all.qwen3-embedding-8b-k5.document.getdoc.sh

Run the GLM-4.7-Flash agent with ColBERTv2 on the three settings:

source pylate/bin/activate
bash ./run_scripts/rq1/GLM-4.7-Flash.queries-all.colbertv2.0-k5.passage.sh
bash ./run_scripts/rq1/GLM-4.7-Flash.queries-all.colbertv2.0-k5.document.sh
bash ./run_scripts/rq1/GLM-4.7-Flash.queries-all.colbertv2.0-k5.document.getdoc.sh

3.2 RQ2: Re-rankers

All scripts assume that the corresponding indices have been built and the agent server is already running.

gpt-oss-20b

Run the gpt-oss-20b agent using ranking pipelines where BM25, SPLADE-v3, or Qwen3-Embedding-8B are used as the retrievers, and monoT5-3B, RankLLaMA-7B or Rank1-7B are used as the re-rankers, on the passage corpus. The re-ranking depth can be modified by changing the RERANKING_DEPTH variable in each script.

source .venv/bin/activate

bash ./run_scripts/rq2/gpt-oss-20b.queries-all.bm25-monot5-3b-msmarco-k5.passage.sh
bash ./run_scripts/rq2/gpt-oss-20b.queries-all.bm25-rankllama-v1-7b-lora-passage-k5.passage.sh

bash ./run_scripts/rq2/gpt-oss-20b.queries-all.splade-v3-monot5-3b-msmarco-k5.passage.sh
bash ./run_scripts/rq2/gpt-oss-20b.queries-all.splade-v3-rankllama-v1-7b-lora-passage-k5.passage.sh

bash ./run_scripts/rq2/gpt-oss-20b.queries-all.qwen3-embedding-8b-monot5-3b-msmarco-k5.passage.sh
bash ./run_scripts/rq2/gpt-oss-20b.queries-all.qwen3-embedding-8b-rankllama-v1-7b-lora-passage-k5.passage.sh

For Rank1-7B, to improve efficiency we host the model on a local server. First launch the Rank1 server:

bash server_rank1-7b.sh

Then run the pipelines with Rank1. Ensure that the RANK1_MODEL_URL variable in each run script matches the Rank1 server endpoint (e.g., http://localhost:8001/v1).

source .venv/bin/activate

bash ./run_scripts/rq2/gpt-oss-20b.queries-all.bm25-rank1-7b-k5.passage.sh
bash ./run_scripts/rq2/gpt-oss-20b.queries-all.splade-v3-rank1-7b-k5.passage.sh
bash ./run_scripts/rq2/gpt-oss-20b.queries-all.qwen3-embedding-8b-rank1-7b-k5.passage.sh

GLM-4.7-Flash

Run the GLM-4.7-Flash agent using ranking pipelines where BM25 or SPLADE-v3 is used as the retriever, and monoT5-3B or RankLLaMA-7B is used as the re-ranker, on the passage corpus. The re-ranking depth can be modified by changing the RERANKING_DEPTH variable in each script.

source .venv/bin/activate

bash ./run_scripts/rq2/GLM-4.7-Flash.queries-all.bm25-monot5-3b-msmarco-k5.passage.sh
bash ./run_scripts/rq2/GLM-4.7-Flash.queries-all.bm25-rankllama-v1-7b-lora-passage-k5.passage.sh

bash ./run_scripts/rq2/GLM-4.7-Flash.queries-all.splade-v3-monot5-3b-msmarco-k5.passage.sh
bash ./run_scripts/rq2/GLM-4.7-Flash.queries-all.splade-v3-rankllama-v1-7b-lora-passage-k5.passage.sh

3.3 RQ3: Query-to-question (Q2Q) reformulation

All scripts assume that the corresponding indices have been built and the agent server is already running.

Run the gpt-oss-20b agent using Q2Q-rewritten queries with BM25, SPLADE-v3, or Qwen3-Embedding-8B as the retriever on the passage corpus:

source .venv/bin/activate

bash ./run_scripts/rq3/gpt-oss-20b.queries-all.q2q-context-bm25-k5.passage.sh
bash ./run_scripts/rq3/gpt-oss-20b.queries-all.q2q-context-splade-v3-k5.passage.sh
bash ./run_scripts/rq3/gpt-oss-20b.queries-all.q2q-context-qwen3-embedding-8b-k5.passage.sh

4. Passage corpus construction

We construct the passage corpus using the same segmentation pipeline employed in previous TREC CAsT tracks.

First, clone the repository of trec-cast-tools, and run the following command to segment documents into passages. The generated passage files will be saved in the ./data/jsonlines directory:

python -u ./corpus_processing/main.py \
  --skip_process_kilt \
  --skip_process_marco_v2 \
  --skip_process_wapo \
  --browsecomp_plus_collection Tevatron/browsecomp-plus-corpus \
  --output_type jsonlines \
  --output_dir ./data/

Next, return to the working directory of this repository and preprocess the generated passage file (e.g., assign passage IDs to each document and compute passage lengths). The following command produces two passage-corpus files: one in the format required by Pyserini and the other in the format required by Tevatron:

python -u passage_corpus_preprocess.py \
--input_dir ./data/jsonlines \
--output_dir ./data/browsecomp-plus-passage

5. Indexing

This section describes how to build retrieval indices for both the passage corpus constructed in this work and the original document corpus. All commands below assume execution from the repository root unless otherwise specified.

5.1 Passage corpus

BM25

source .venv/bin/activate

python -m pyserini.index.lucene \
--collection JsonCollection \
--input ./data/browsecomp-plus-passage/browsecomp-plus-passage-pyserini/ \
--index ./indexes/index.bm25.passage \
--generator DefaultLuceneDocumentGenerator \
--threads 16 \
--storePositions --storeDocvectors --storeRaw

SPLADE-v3

Regarding SPLADE-v3, first ensure that Tevatron is installed, and switch to the Tevatron directory. Run the following commands to encode the passage corpus:

source .venv/bin/activate

mkdir -p ./indexes/encoding.splade-v3.passage

CUDA_DEVICE_ORDER=PCI_BUS_ID CUDA_VISIBLE_DEVICES=0 \
python ./examples/splade/encode_splade.py \
  --model_name_or_path naver/splade-v3 \
  --tokenizer_name bert-base-uncased \
  --fp16 \
  --passage_max_len 512 \
  --per_device_eval_batch_size 128 \
  --dataset_path ./data/browsecomp-plus-passage/browsecomp-plus-passage-tevatron.jsonl \
  --dataset_config jsonl \
  --output_dir ./indexes/encoding.splade-v3.passage \
  --encode_output_path ./indexes/encoding.splade-v3.passage/corpus.jsonl

After encoding, build the Lucene index using Pyserini:

python -m pyserini.index.lucene \
  --collection JsonVectorCollection \
  --input ./indexes/encoding.splade-v3.passage/ \
  --index ./indexes/index.splade-v3.passage \
  --generator DefaultLuceneDocumentGenerator \
  --threads 16 \
  --impact --pretokenized

RepLLaMA

source .venv/bin/activate

mkdir ./indexes/index.repllama-v1-7b-lora-passage.passage
CUDA_DEVICE_ORDER=PCI_BUS_ID CUDA_VISIBLE_DEVICES=0 \
python -m tevatron.retriever.driver.encode \
  --output_dir=./temp \
  --model_name_or_path meta-llama/Llama-2-7b-hf \
  --lora_name_or_path castorini/repllama-v1-7b-lora-passage \
  --lora \
  --query_prefix "query: " \
  --passage_prefix "passage: " \
  --bf16 \
  --pooling eos \
  --append_eos_token \
  --normalize \
  --per_device_eval_batch_size 32 \
  --query_max_len 512 \
  --passage_max_len 512 \
  --dataset_path ./data/browsecomp-plus-passage/browsecomp-plus-passage-tevatron.jsonl \
  --dataset_config jsonl \
  --encode_output_path ./indexes/index.repllama-v1-7b-lora-passage.passage/corpus.pkl 

Qwen3-Embedding-8B

source .venv/bin/activate

mkdir ./indexes/index.qwen3-embedding-8b.passage
CUDA_DEVICE_ORDER=PCI_BUS_ID CUDA_VISIBLE_DEVICES=0 \
python -m tevatron.retriever.driver.encode \
  --model_name_or_path Qwen/Qwen3-Embedding-8B \
  --dataset_path ./data/browsecomp-plus-passage/browsecomp-plus-passage-tevatron.jsonl \
  --dataset_config jsonl \
  --encode_output_path ./indexes/index.qwen3-embedding-8b.passage/corpus.pkl \
  --passage_max_len 512 \
  --normalize \
  --pooling eos \
  --passage_prefix "" \
  --per_device_eval_batch_size 64 \
  --fp16

ColBERTv2

source pylate/bin/activate

CUDA_DEVICE_ORDER=PCI_BUS_ID CUDA_VISIBLE_DEVICES=0 \
 python colbert.py \
--model_name colbert-ir/colbertv2.0 \
--corpus_path ./data/browsecomp-plus-passage/browsecomp-plus-passage-tevatron.jsonl \
--index_folder ./indexes/index.colbertv2.0-d512.passage \
--batch_size 2048 \
--max_seq_length 512 \
--document_length 512 \
--query_length 512 \
--chunk_size 100000 --shard_count 4 \
--build_index 

5.2 Document corpus

We use the pre-built BM25 index and the Qwen3-Embedding-8B index for the document corpus released by the BrowseComp-Plus authors. Here we show how to index the document corpus using SPLADE-v3, RepLLaMA and ColBERTv2.

SPLADE-v3

source .venv/bin/activate
# First ensure that Tevatron is installed and switch to the Tevatron directory.
python ./examples/splade/encode_splade.py \
  --model_name_or_path naver/splade-v3 \
  --tokenizer_name bert-base-uncased \
  --fp16 \
  --passage_max_len 512 \
  --per_device_eval_batch_size 128 \
  --dataset_name Tevatron/browsecomp-plus-corpus \
  --output_dir ./indexes/index.splade-v3.document \
  --encode_output_path ./indexes/encoding.splade-v3.document/corpus.jsonl

python -m pyserini.index.lucene \
  --collection JsonVectorCollection \
  --input ./indexes/encoding.splade-v3.document/ \
  --index ./indexes/index.splade-v3.document \
  --generator DefaultLuceneDocumentGenerator \
  --threads 16 \
  --impact --pretokenized

RepLLaMA

source .venv/bin/activate

mkdir ./indexes/index.repllama-v1-7b-lora-passage.document

CUDA_DEVICE_ORDER=PCI_BUS_ID CUDA_VISIBLE_DEVICES=0 \
python -m tevatron.retriever.driver.encode \
  --output_dir=./temp \
  --model_name_or_path meta-llama/Llama-2-7b-hf \
  --lora_name_or_path castorini/repllama-v1-7b-lora-passage \
  --lora \
  --query_prefix "query: " \
  --passage_prefix "passage: " \
  --bf16 \
  --pooling eos \
  --append_eos_token \
  --normalize \
  --per_device_eval_batch_size 64 \
  --query_max_len 512 \
  --passage_max_len 512 \
  --dataset_name Tevatron/browsecomp-plus-corpus \
  --encode_output_path ./indexes/index.repllama-v1-7b-lora-passage.document/corpus.pkl 

ColBERTv2

source pylate/bin/activate

CUDA_DEVICE_ORDER=PCI_BUS_ID CUDA_VISIBLE_DEVICES=0 \
 python colbert.py \
--model_name colbert-ir/colbertv2.0 \
--corpus_path Tevatron/browsecomp-plus-corpus \
--index_folder ./indexes/index.colbertv2.0-d512.document \
--batch_size 2048 \
--max_seq_length 512 \
--document_length 512 \
--query_length 512 \
--shard_count 4 \
--build_index 

6. Decrypting encrypted runs

Download the encrypted run files from here and decrypt them:

bash ./scripts_encrypt_decrypt_run/download_runs.sh
bash ./scripts_encrypt_decrypt_run/decrypt_runs.sh

This script will decrypt all run files locally and generate the corresponding plaintext execution traces.

7. Contact

If you have any questions or suggestions, please contact us at:

• Chuan Meng: chuan.meng@ed.ac.uk
• Litu Ou: litu.ou@ed.ac.uk