vkola-lab/ClinicalTrials.gov

Special Notice

For this dataset, we open-source our processed clinical trial records from ClinicalTrials.gov, filtering for studies that had completed recruitment and were classified as Phase 3 or Phase 4, or that investigated device-based or behavioral interventions. This selection yielded $156,887$ trials as of March 2025.

Agentic Medical RAG

Agentic Memory-augmented Retrieval and Evidence Grounding in Medicine

🎉 Announcements

[2026.06.01] Our Agentic Medical RAG manuscript is published in International Journal of Medical Informatics. Link to the paper and Supplementary Material.

📚 Table of contents

• Agentic medical RAG
• Installation
• Quick start
  • Retrieval
  • Reranking
  • Generation
• Prompt format
  • System prompt
  • User prompt for multiple-choice QA
  • User prompt for open-ended QA
  • User prompt for cache-and-prune memory bank mechanism
• Toolkit description
• Dataset description
• Benchmarks and results
  • Benchmark datasets
  • Multiple-choice QA performance
  • Open-ended QA performance
  • Tool-usage analysis
  • Ablation study
• Real-world deployment
• Algorithm
• Structure of the code
• Citation
• Contact
• Contributions
• Acknowledgements

Agentic medical RAG

We developed a unified, open-source LLM-based agentic system that integrates document retrieval, re-ranking, evidence grounding, and diagnosis generation to support dynamic, multi-step medical reasoning. Our system features a lightweight retrieval-augmented generation pipeline coupled with a cache-and-prune memory bank, enabling efficient long-context inference beyond standard LLM limits. The system autonomously invokes specialized tools, eliminating the need for manual prompt engineering or brittle multi-stage templates.

💻 Installation

At the time, this work was conducted using 4 NVIDIA L40S GPUs. To facilitate package installation and dependency configuration, we validated our agentic system on 2 NVIDIA RTX PRO 6000 Blackwell GPUs running CUDA 13.2 and 0.17.0+cu132.

Please NOTE that our vLLM installation is built from source instead of direct pip install.

First, create a Python 3.10 environment (higher versions should work as well), then

pip install -r requirements.txt

Then, the most important part is to install/compile vLLM:

• Manually download the vLLM 0.17.0 source code from https://github.com/vllm-project/vllm/releases/tag/v0.17.0.
• tar -xvzf vllm-0.17.0.tar.gz
• pip install --editable vllm-0.17.0 --no-cache-dir

Then, the vLLM will be compiled from source and installed. Other vLLM versions (e.g., 0.6.3), can also be compiled in this way.

Running demonstration (left tmux serves the model, the other tmux runs the Python/Shell script)

To set up the Transformers and Triton saving path, run the following commands in your terminal:

For example, if your path is /projectnb/vkolagrp/:

export TRANSFORMERS_CACHE=/projectnb/vkolagrp/.cache
export HF_HOME=/projectnb/vkolagrp/.cache
export HF_DATASETS_CACHE=/projectnb/vkolagrp/.cache
export TRITON_CACHE_DIR=/projectnb/vkolagrp/.cache/.triton

Please change /projectnb/vkolagrp/ to your own path. After running the commands, all your Transformers models and datasets will be saved in the paths you defined.

🚀 Quick start

We use a two-stage retrieval pipeline:

1. Document retrieval: SPECTER performs coarse-grained semantic search over each evidence source and retrieves the top-K candidate documents.
2. Evidence reranking: gte-Qwen2-7B-instruct computes fine-grained query–document similarity and selects the top-R most relevant documents for downstream evidence grounding and medical reasoning.

By default, the system retrieves $32$ candidates from each of the six sources, producing $192$ candidates in total, and reranks them to retain the $32$ most relevant documents. This multi-source design provides complementary evidence from biomedical literature, clinical studies, textbooks, case reports, and general medical knowledge.

Retrieval

The retrieval module identifies clinically relevant evidence from six medical and scientific sources:

• PubMed abstracts
• PubMed Central full-text articles
• Medical textbooks
• ClinicalTrials.gov records
• Wikipedia articles
• New England Journal of Medicine (NEJM) clinical case reports

  Due to copyright restrictions, the NEJM clinical case report dataset is not included in this open-source release.

To retrieve documents for all supported benchmarks, run:

sh run_retrieve.sh

To retrieve documents for a specific benchmark, run the corresponding command:

python retrieve.py --benchmark USMLE_STEP_1 --save_folder "RETRIEVED_USMLE_STEP_1"
python retrieve.py --benchmark USMLE_STEP_2 --save_folder "RETRIEVED_USMLE_STEP_2"
python retrieve.py --benchmark USMLE_STEP_3 --save_folder "RETRIEVED_USMLE_STEP_3"
python retrieve.py --benchmark MedQA --save_folder "RETRIEVED_MedQA"
python retrieve.py --benchmark MedExpQA --save_folder "RETRIEVED_MedExpQA"
python retrieve.py --benchmark PubMedQA --save_folder "RETRIEVED_PubMedQA"

Reranking

To rerank the retrieved documents for all supported benchmarks, run:

sh run_rerank.sh

To rerank documents for a specific benchmark, run the corresponding command:

python rerank.py --benchmark USMLE_STEP_1 --doc_folder "RETRIEVED_USMLE_STEP_1" --save_folder "RERANKED_USMLE_STEP_1"
python rerank.py --benchmark USMLE_STEP_2 --doc_folder "RETRIEVED_USMLE_STEP_2" --save_folder "RERANKED_USMLE_STEP_2"
python rerank.py --benchmark USMLE_STEP_3 --doc_folder "RETRIEVED_USMLE_STEP_3" --save_folder "RERANKED_USMLE_STEP_3"
python rerank.py --benchmark MedQA --doc_folder "RETRIEVED_MedQA" --save_folder "RERANKED_MedQA"
python rerank.py --benchmark MedExpQA --doc_folder "RETRIEVED_MedExpQA" --save_folder "RERANKED_MedExpQA"
python rerank.py --benchmark PubMedQA --doc_folder "RETRIEVED_PubMedQA" --save_folder "RERANKED_PubMedQA"

Generation

sh run_diagnosis.sh

or

# USMLE STEP 3 Benchmark
python diagnosis.py --benchmark USMLE_STEP_3 --mode "tool_using_open_book" --doc_folder "RERANKED_USMLE_STEP_3" --save_folder "RESULTS_USMLE_STEP_3"
python diagnosis.py --benchmark USMLE_STEP_3 --mode "tool_using_close_book" --doc_folder "RERANKED_USMLE_STEP_3" --save_folder "RESULTS_USMLE_STEP_3"
python diagnosis.py --benchmark USMLE_STEP_3 --mode "direct_prompt" --doc_folder "RERANKED_USMLE_STEP_3" --save_folder "RESULTS_USMLE_STEP_3"
python diagnosis.py --benchmark USMLE_STEP_3 --mode "direct_prompt_with_evidence" --doc_folder "RERANKED_USMLE_STEP_3" --save_folder "RESULTS_USMLE_STEP_3"
python diagnosis.py --benchmark USMLE_STEP_3 --mode "unstructured_prompt_with_evidence" --doc_folder "RERANKED_USMLE_STEP_3" --save_folder "RESULTS_USMLE_STEP_3"

# USMLE STEP 2 Benchmark
python diagnosis.py --benchmark USMLE_STEP_2 --mode "tool_using_open_book" --doc_folder "RERANKED_USMLE_STEP_2" --save_folder "RESULTS_USMLE_STEP_2"
python diagnosis.py --benchmark USMLE_STEP_2 --mode "tool_using_close_book" --doc_folder "RERANKED_USMLE_STEP_2" --save_folder "RESULTS_USMLE_STEP_2"
python diagnosis.py --benchmark USMLE_STEP_2 --mode "direct_prompt" --doc_folder "RERANKED_USMLE_STEP_2" --save_folder "RESULTS_USMLE_STEP_2"
python diagnosis.py --benchmark USMLE_STEP_2 --mode "direct_prompt_with_evidence" --doc_folder "RERANKED_USMLE_STEP_2" --save_folder "RESULTS_USMLE_STEP_2"
python diagnosis.py --benchmark USMLE_STEP_2 --mode "unstructured_prompt_with_evidence" --doc_folder "RERANKED_USMLE_STEP_2" --save_folder "RESULTS_USMLE_STEP_2"

# USMLE STEP 1 Benchmark
python diagnosis.py --benchmark USMLE_STEP_1 --mode "tool_using_open_book" --doc_folder "RERANKED_USMLE_STEP_1" --save_folder "RESULTS_USMLE_STEP_1"
python diagnosis.py --benchmark USMLE_STEP_1 --mode "tool_using_close_book" --doc_folder "RERANKED_USMLE_STEP_1" --save_folder "RESULTS_USMLE_STEP_1"
python diagnosis.py --benchmark USMLE_STEP_1 --mode "direct_prompt" --doc_folder "RERANKED_USMLE_STEP_1" --save_folder "RESULTS_USMLE_STEP_1"
python diagnosis.py --benchmark USMLE_STEP_1 --mode "direct_prompt_with_evidence" --doc_folder "RERANKED_USMLE_STEP_1" --save_folder "RESULTS_USMLE_STEP_1"
python diagnosis.py --benchmark USMLE_STEP_1 --mode "unstructured_prompt_with_evidence" --doc_folder "RERANKED_USMLE_STEP_1" --save_folder "RESULTS_USMLE_STEP_1"

# PubMedQA Benchmark
python diagnosis.py --benchmark PubMedQA --mode "tool_using_open_book" --doc_folder "RERANKED_PubMedQA" --save_folder "RESULTS_PubMedQA"
python diagnosis.py --benchmark PubMedQA --mode "tool_using_close_book" --doc_folder "RERANKED_PubMedQA" --save_folder "RESULTS_PubMedQA"
python diagnosis.py --benchmark PubMedQA --mode "direct_prompt" --doc_folder "RERANKED_PubMedQA" --save_folder "RESULTS_PubMedQA"
python diagnosis.py --benchmark PubMedQA --mode "direct_prompt_with_evidence" --doc_folder "RERANKED_PubMedQA" --save_folder "RESULTS_PubMedQA"
python diagnosis.py --benchmark PubMedQA --mode "unstructured_prompt_with_evidence" --doc_folder "RERANKED_PubMedQA" --save_folder "RESULTS_PubMedQA"

# MedExpQA Benchmark
python diagnosis.py --benchmark MedExpQA --mode "tool_using_open_book" --doc_folder "RERANKED_MedExpQA" --save_folder "RESULTS_MedExpQA"
python diagnosis.py --benchmark MedExpQA --mode "tool_using_close_book" --doc_folder "RERANKED_MedExpQA" --save_folder "RESULTS_MedExpQA"
python diagnosis.py --benchmark MedExpQA --mode "direct_prompt" --doc_folder "RERANKED_MedExpQA" --save_folder "RESULTS_MedExpQA"
python diagnosis.py --benchmark MedExpQA --mode "direct_prompt_with_evidence" --doc_folder "RERANKED_MedExpQA" --save_folder "RESULTS_MedExpQA"
python diagnosis.py --benchmark MedExpQA --mode "unstructured_prompt_with_evidence" --doc_folder "RERANKED_MedExpQA" --save_folder "RESULTS_MedExpQA"

# MedQA Benchmark
python diagnosis.py --benchmark MedQA --mode "tool_using_open_book" --doc_folder "RERANKED_MedQA" --save_folder "RESULTS_MedQA"
python diagnosis.py --benchmark MedQA --mode "tool_using_close_book" --doc_folder "RERANKED_MedQA" --save_folder "RESULTS_MedQA"
python diagnosis.py --benchmark MedQA --mode "direct_prompt" --doc_folder "RERANKED_MedQA" --save_folder "RESULTS_MedQA"
python diagnosis.py --benchmark MedQA --mode "direct_prompt_with_evidence" --doc_folder "RERANKED_MedQA" --save_folder "RESULTS_MedQA"
python diagnosis.py --benchmark MedQA --mode "unstructured_prompt_with_evidence" --doc_folder "RERANKED_MedQA" --save_folder "RESULTS_MedQA"

📜 Prompt format

System prompt

system_prompt = (
  "You are a medical professional specializing in evidence-based medicine (EBM). "
  "Your role is to answer questions using a systematic approach, integrating the best available "
  "research evidence, clinical expertise, and patient-specific factors."
)

User prompt for multiple-choice QA

tool_evidence_prompt = """Here is the background information and question about the patient:
<background>
{background}
</background>

The available answer options are:
<option>
{option}
</option>

Follow these steps to answer the question:
1. Compare each option with the case details, analyzing key clues in the text to identify the best choice.
2. If the question can be answered through comparison, directly return the best option term with the option capital within <final_result></final_result> tags, placing the explanation outside of the <final_result> tags.
3. If multiple options are plausible or additional evidence is needed for better decision-making, enable search to find credible sources.
4. Analyze the relevance between each document and the patient's presentation, followed by a systematic search to locate relevant evidence applicable to the patient’s case.
5. While we are continuing to provide additional evidence, iterate the previous step to analyze more additional evidence.
6. Once sufficient information is gathered, return the best option term with the option capital within <final_result></final_result> tags, placing the explanation outside of the <final_result> tags."""

User prompt for open-ended QA

tool_evidence_prompt = """Here is the background information and question about the patient:
<background>
{background}
</background>

Follow these steps to answer the question:
1. Compare each option with the case details, analyzing key clues in the text to identify the best choice.
2. If the question can be answered through comparison, directly return the best option term with the option capital within <final_result></final_result> tags, placing the explanation outside of the <final_result> tags.
3. If multiple options are plausible or additional evidence is needed for better decision-making, enable search to find credible sources.
4. Analyze the relevance between each document and the patient's presentation, followed by a systematic search to locate relevant evidence applicable to the patient’s case.
5. While we are continuing to provide additional evidence, iterate the previous step to analyze more additional evidence.
6. Once sufficient information is gathered, return the best option term with the option capital within <final_result></final_result> tags, placing the explanation outside of the <final_result> tags."""

User prompt for cache-and-prune memory bank mechanism

document_revise_template = """Here are the selected relevant documents related to the patient's care:
<document>
{document}
</document>

Review your answer and return the best option term with the option capital within the <final_result></final_result> tags, leave the explanation outside of the <final_result> tags."""

🧰 Toolkit description

we designed five specialized tools to serve as the diagnostic aid kit within our AI agent. For question interpretation, the agent uses perform_comparison to handle multiple choice tasks and generate_options for open-ended scenarios, enabling flexible reasoning formats. In particular, generate_options is tailored for scenarios lacking predefined choices, enabling the agent to propose plausible answer candidates from the contextual details of the case. Additionally, the enable_search tool determines whether external evidence is necessary to support a diagnosis. To facilitate evidence retrieval and interpretation, relevance_analysis evaluates the semantic alignment between the patient’s case and retrieved documents, while locate_evidence identifies and grounds specific articles most pertinent to the diagnosis.

📚 Dataset description

The corpus includes research articles published under Creative Commons licenses from leading biomedical journals indexed in PubMed Central. We also incorporated clinical trial records from ClinicalTrials.gov, filtering for studies that had completed recruitment and were classified as Phase 3 or Phase 4, or that investigated device-based or behavioral interventions. This selection yielded $156,887$ trials as of March 2025. To enhance real-world clinical applicability, we included $1,479$ clinical case reports published by the New England Journal of Medicine (NEJM) between 2016 and March2025. We further adopted pre-indexed corpora of PubMed abstracts and Wikipedia entries from Xiong et al., which have demonstrated strong utility for medical question answering (QA) tasks. Finally, we leveraged $8,226$ open-access medical textbooks from the NLM LitArch Open Access Subset, hosted by the U.S. National Library of Medicine.

🏆 Benchmarks and results

Benchmark datasets

We evaluated our system using five medical question answering benchmarks: the United States medical licensing examination (USMLE) Steps 1-3, free-form multiple-choice open-domain medical QA (MedQA), and medical explanation-based QA (MedExpQA) (Table 3). Each benchmark includes clinical case descriptions, multiple-choice options, and a correct answer.

Multiple-choice QA performance

We manually evaluated the accuracy of all models, including BioMistral (7B), OpenBioLLM (8B and 70B), UltraMedical (8B and 70B), and PodGPT (70B).

Open-ended QA performance

We manually evaluated the performance of all models, including BioMistral (7B), OpenBioLLM (8B and 70B), UltraMedical (8B and 70B), and PodGPT (70B), using three embedding-based metrics: BERTScore F1 computed with deberta-xlarge-mnli, similarity scores from SFR-Embedding-2_R, and similarity scores from gte-Qwen2-7B-instruct models. We provide evaluation code here.

Tool-usage analysis

We analyzed the agent's tool usage patterns, and provide our code (Figure 2(a)) and code (Figure 2(b)) to generate these figures.

Ablation study

We compared performance with and without tool access to evaluate the impact of incorporating tools into the agentic pipeline.

🔥 Real-world deployment

For real-world deployment, please refer to the vLLM Distributed Inference and Serving and OpenAI Compatible Server. We provide a deployment script here.

At the time, the vLLM version we were using was 0.6.3 to serve the Qwen2.5 72B model. Please check this version.

vLLM can be deployed as a server that implements the OpenAI API protocol. This allows vLLM to be used as a drop-in replacement for applications using OpenAI API. By default, it starts the server at http://localhost:8000.

VLLM_RPC_TIMEOUT=100000 vllm serve Qwen/Qwen2.5-72B-Instruct \
    --trust-remote-code \
    --dtype bfloat16 \
    --max-model-len 32768 \
    --tensor_parallel_size 4 \
    --gpu-memory-utilization 0.95 \
    --api-key token-abc123 \
    --enable-auto-tool-choice \
    --tool-call-parser hermes

Please check here if you wanna change Engine Arguments.

After you download the Qwen model, you need to change the default prompt in the tokenizer_config.json file (this line) by changing the You may call one or more functions to assist with the user query. to You should only call one tool at a time to assist with the user query..

The commands used to serve the open-source scientific and biomedical language models evaluated in this work are provided below:

Model: BioMistral-7B
Link: https://huggingface.co/BioMistral/BioMistral-7B
VLLM_RPC_TIMEOUT=10000 vllm serve BioMistral/BioMistral-7B \
    --trust-remote-code \
    --dtype bfloat16 \
    --max-model-len 2048 \
    --distributed-executor-backend mp \
    --tensor_parallel_size 1 \
    --gpu-memory-utilization 0.95 \
    --api-key token-abc123 \
    --max-num-seqs 1

Model: aaditya/Llama3-OpenBioLLM-8B
Link: https://huggingface.co/aaditya/Llama3-OpenBioLLM-8B
VLLM_RPC_TIMEOUT=10000 vllm serve aaditya/Llama3-OpenBioLLM-8B \
    --trust-remote-code \
    --dtype bfloat16 \
    --max-model-len 8192 \
    --distributed-executor-backend mp \
    --tensor_parallel_size 1 \
    --gpu-memory-utilization 0.95 \
    --api-key token-abc123 \
    --tokenizer meta-llama/Meta-Llama-3-8B-Instruct

Model: aaditya/Llama3-OpenBioLLM-70B
Link: https://huggingface.co/aaditya/Llama3-OpenBioLLM-70B
VLLM_RPC_TIMEOUT=10000 vllm serve aaditya/Llama3-OpenBioLLM-70B \
    --trust-remote-code \
    --dtype bfloat16 \
    --max-model-len 8192 \
    --distributed-executor-backend mp \
    --tensor_parallel_size 4 \
    --gpu-memory-utilization 0.95 \
    --api-key token-abc123 \
    --tokenizer meta-llama/Meta-Llama-3-70B-Instruct

Model: TsinghuaC3I/Llama-3.1-8B-UltraMedical
Link: https://huggingface.co/TsinghuaC3I/Llama-3.1-8B-UltraMedical
VLLM_RPC_TIMEOUT=10000 vllm serve TsinghuaC3I/Llama-3.1-8B-UltraMedical \
    --trust-remote-code \
    --dtype bfloat16 \
    --max-model-len 8192 \
    --distributed-executor-backend mp \
    --tensor_parallel_size 1 \
    --gpu-memory-utilization 0.95 \
    --api-key token-abc123 \
    --max-num-seqs 1

Model: TsinghuaC3I/Llama-3-70B-UltraMedical
Link: https://huggingface.co/TsinghuaC3I/Llama-3-70B-UltraMedical
VLLM_RPC_TIMEOUT=10000 vllm serve TsinghuaC3I/Llama-3-70B-UltraMedical \
    --trust-remote-code \
    --dtype bfloat16 \
    --max-model-len 8192 \
    --distributed-executor-backend mp \
    --tensor_parallel_size 4 \
    --gpu-memory-utilization 0.95 \
    --api-key token-abc123

Model: PodGPT 70B
Link: https://huggingface.co/shuyuej/Llama-3.3-70B-Instruct-2048/tree/main/checkpoint-18640
VLLM_RPC_TIMEOUT=10000 vllm serve meta-llama/Llama-3.3-70B-Instruct \
    --trust-remote-code \
    --dtype bfloat16 \
    --max-model-len 8192 \
    --distributed-executor-backend mp \
    --tensor_parallel_size 4 \
    --gpu-memory-utilization 0.95 \
    --api-key token-abc123 \
    --enable-lora \
    --lora-modules adapter=Llama-3.3-70B-Instruct-2048/checkpoint-18640

🧐 Algorithm

Given a patient’s background and clinical question, the AI agent dynamically determines whether external evidence is required and autonomously invokes the appropriate tools.

When evidence retrieval is enabled, the system:

1. Retrieves the top-K candidate documents from the medical evidence database.
2. Reranks the candidates and selects the top-R most relevant documents.
3. Processes the reranked documents incrementally in batches.
4. Identifies and grounds relevant evidence from each batch.
5. Stores useful evidence in a cache-and-prune memory bank while removing irrelevant information from the active context.
6. Continues searching and grounding until sufficient evidence has been collected.
7. Generates the final answer using the patient information, reasoning history, and grounded evidence retained in memory.

This iterative design allows the agent to integrate evidence beyond the model’s effective context-window limit while ensuring that its final response is informed by the most relevant medical information.

🖼️ Structure of the code

At the root of the project, you will see:

.
├── config.yml
├── config_benchmark.yml
├── retrieve.py
├── rerank.py
├── diagnosis.py
├── run_retrieve.sh
├── run_rerank.sh
├── run_diagnosis.sh
├── lib
│   ├── framework.py
│   ├── prompts.py
│   └── tools.py
├── database
│   └── rag.py
├── benchmarks
│   ├── MedExpQA
│   ├── MedQA
│   ├── PubMedQA
│   └── USMLE
├── evaluation-open-ended-QA
│   ├── bert_score.py
│   ├── GTE-embedding-Model.py
│   └── SFR-embedding-Model.py
├── results
│   ├── Ablation-study
│   ├── Multiple-choice-QA
│   ├── Open-ended-QA
│   └── Qwen3-models-for-rebuttal
└── utils
    ├── answer_extraction.py
    ├── benchmark_utils.py
    └── utils.py

🙏 Citation

If you find our work useful in your research, please consider citing it in your publications. We provide a BibTeX entry below.

@article{Jia2026agenticsystem,
    title = {Agentic memory-augmented retrieval and evidence grounding for medical question-answering tasks},
    author = {Shuyue Jia and Subhrangshu Bit and Varuna H. Jasodanand and Yi Liu and Vijaya B. Kolachalama},
    journal = {International Journal of Medical Informatics},
    volume = {212},
    pages = {106339},
    year = {2026},
    issn = {1386-5056},
    doi = {https://doi.org/10.1016/j.ijmedinf.2026.106339},
    url = {https://www.sciencedirect.com/science/article/pii/S1386505626000791},
}

📧 Contact

Core Contributor and Maintainer:

• Shuyue Jia

Other Contributor and Maintainer:

• Subhrangshu Bit
• Varuna H. Jasodanand
• Yi Liu

If you have any questions, please drop us an email at brucejia@bu.edu.

🔨 Contributions

We always welcome contributions to help make our Agentic Medical RAG system better. If you would like to contribute, please submit a pull request.

🙌 Acknowledgements

This repository is maintained by members of the Kolachalama Laboratory in Boston, MA.

⬆️ Back to Top