Patients with cancer increasingly seek herbs and dietary supplements for symptom relief [1]. However, concurrent use of supplements with cancer treatment can lead to harmful adverse events and herb-drug interactions [2]. Additionally, many oncology providers have limited knowledge/training to counsel patients [3]. There is also a growing amount of misinformation about cancer treatments and exaggerated benefits of natural products for improving outcomes including curing cancer [4].

The About Herbs app—developed by the Integrative Medicine Service at Memorial Sloan Kettering Cancer Center (MSK)—delivers content from the About Herbs database that contains the latest research on 317 entries including herbs, minerals, and vitamins [5]. Currently, the app uses a keyword-based search system, which is inadequate to answer complex queries typically asked by providers. This can hinder their communication with patients on this important topic.

Since 2022, large language models (LLMs) are being increasingly used to obtain health-related information. While LLMs outperform online searches, they are prone to producing inaccurate responses or hallucinations [6], which can be especially harmful in oncology settings. Retrieval-augmented generation (RAG) is a technique that can improve the accuracy of responses by integrating the LLM with a searchable knowledge base [7]. To enhance the search functionality of the About Herbs app, we developed and evaluated a RAG-based LLM grounded in the About Herbs database.

In total, we evaluated 1300 question-answer pairs across 5 expert-curated questions, 26 monographs, and 9 LLMs and AboutHerbsAI. Interrater agreement was substantial for across rater pairs (κ=0.66‐0.74). Our model was superior to the 9 LLMs in both correctness and conciseness of responses with no significant differences in completeness (Figure 1).

This study found that a RAG-based LLM substantially improved correctness and conciseness of responses to common questions about herbs and dietary supplements, when compared with 9 general-purpose LLMs, but slightly underperformed in completeness. We consider this acceptable given the benefit of grounding responses in an evidence-based database rather than open internet sources, which are prone to misinformation.

These results align with prior concerns about LLM hallucinations [6] and complementary and alternative medicine misinformation [4] and support our approach as a proof-of-concept study for integrating evidence-grounded LLM assistance [7] to improve the quality and accuracy of responses. This study is timely as oncology care is becoming increasingly digitized. An important limitation is the confinement of our model to a curated database that limits response breadth, especially when newer evidence is not yet included. Second, its performance was assessed in a controlled setting with limited content and predefined questions. Therefore, generalizability, and real-world queries—particularly those using vague or nonmedical/nontechnical language (eg, “hangover”)—may pose challenges [10].

Future work will include system enhancements and evaluation of the clinical/patient-facing impacts of About Herbs. Database curation will involve regular ingestion of the latest literature to ensure timeliness. To support patients with varying health literacy needs, the app will offer a configurable layperson-oriented version to deliver information. Potential clinical implementations include embedding the RAG-based LLM within the About Herbs app to help physicians deliver evidence-based recommendations to minimize risks associated with the concomitant use of supplements during cancer treatment. Evaluation will involve assessing clinical significance, including effects on herb-related counseling, documentation, and safety, alongside integration into electronic medical records and clinic workflows.