Towards Doctor-Like Reasoning: Medical RAG Fusing Knowledge with Patient Analogy through Textual Gradients

DoctorRAG is an advanced medical Retrieval-Augmented Generation (RAG) framework that emulates doctor-like reasoning by integrating explicit medical knowledge with experiential insights from real patient cases. Med-TextGrad is a multi-agent, iterative answer refinement process that further enhances the accuracy, relevance, and safety of generated responses.

Figure

Figure: Overview of the DoctorRAG framework, illustrating the integration of medical knowledge and patient analogical reasoning via Med-TextGrad.

---

Features

• Hybrid Retrieval: Combines knowledge base (medical expertise) and patient base (clinical experience) for context-rich answers.
• Concept Tagging & Declarative Transformation: Structures both queries and knowledge for precise, concept-aware retrieval.
• Iterative Answer Optimization: Med-TextGrad uses multi-agent textual gradients to iteratively refine answers.
• Multilingual & Multitask: Supports Chinese, English, and French datasets for diagnosis, QA, treatment recommendation, and text generation.
• Plug-and-Play LLMs: Compatible with OpenAI, DeepSeek, and other major LLM APIs.
• Efficient Retrieval: Uses FAISS for fast, scalable similarity search.

---

Project Structure

DoctorRAG/
├── Datasets/           # Raw and processed datasets (CSV, by benchmark/language)
├── Knowledge_Base/     # Structured medical knowledge bases (multi-language, FAISS indices)
├── Patient_Base/       # Patient case data and FAISS indices
├── Outputs/            # Results and logs from experiments
├── Scripts/            # All experiment, training, and evaluation code
│   ├── DoctorRAG/      # Main DoctorRAG pipeline scripts
│   └── Med-TextGrad/   # Med-TextGrad iterative refinement and evaluation
├── Utils/              # Utility scripts (concept tagging, declarative transformation, etc.)
├── requirements.txt    # Python dependencies

---

Getting Started

Prerequisites

• Python 3.8+
• API access for OpenAI, DeepSeek, or other supported LLMs
• Sufficient disk space for datasets and FAISS indices

Installation

1. Clone the repository:

   git clone [<https://github.com/YuxingLu613/DoctorRAG>](https://github.com/YuxingLu613/DoctorRAG)
   cd DoctorRAG

2. Install dependencies:

   pip install -r requirements.txt

3. Set up API keys:

   • Configure your OpenAI/DeepSeek API keys as required in the scripts (see comments in Scripts/DoctorRAG/ and Scripts/Med-TextGrad/).

---

Usage

1. Data Preparation

• Place your datasets in Datasets/ and knowledge/patient bases in their respective folders.
• Each dataset/knowledge base should be organized by language and benchmark (see folder structure above).

2. Running DoctorRAG

• Use the scripts in Scripts/DoctorRAG/ to run retrieval-augmented generation on your chosen dataset.
• Example:

  python Scripts/DoctorRAG/DDXPlus_EN_DD.py

3. Iterative Answer Refinement (Med-TextGrad)

• Use Scripts/Med-TextGrad/Med_TextGrad.py to iteratively refine answers.
• For pairwise evaluation, use Scripts/Med-TextGrad/Pairwise-Rater.py.

4. Outputs

• Results and logs are saved in Outputs/ under the relevant method and experiment.

---

Utilities

• Utils/MedQA_Declarative_Sentence_*.py: Convert knowledge chunks to declarative statements.
• Utils/MedQA_Concept_Tagging_*.py: Tag statements with medical concepts.
• Utils/Knowledge_Base_Faiss.py: Build and query FAISS indices for fast retrieval.

---

Datasets

• Datasets: Multilingual, multitask medical QA and patient case datasets (see Datasets/).
• Knowledge Bases: Structured medical knowledge in English, Chinese, and French (see Knowledge_Base/).
• Patient Bases: De-identified patient records for experience-based retrieval (see Patient_Base/).

Note: Due to size restrictions, some folders may contain placeholders.

---

Citation

If you use this code or data, please cite our paper (TBA):

---

Contact

For questions or collaborations, please contact [[email protected]].