Med.ai ASK: an agentic system for biomedical question answering.

Area of Science:

• Biomedical Informatics
• Artificial Intelligence in Science
• Natural Language Processing

Background:

• Generative AI is revolutionizing scientific knowledge access.
• Current AI solutions struggle with the nuance of biomedical research questions.
• There is a need for reliable, grounded AI-powered biomedical information retrieval.

Purpose of the Study:

• To develop Med.ai ASK, an agentic question-answering system for biomedical inquiries.
• To enhance accuracy and reliability in AI-generated biomedical responses.
• To create a system that parses complex research questions effectively.

Main Methods:

• Utilized the ReAct framework and Self-Discover for tool-calling and reasoning.
• Integrated multiple biomedical knowledge bases, vector databases, APIs, and NER tools.
• Ingested 44 million biomedical documents and evaluated on diverse QA datasets.

Main Results:

• Med.ai ASK demonstrates strong performance, accuracy, and reduced hallucinations compared to other AI solutions.
• Human and LLM evaluations show aligned assessments, supporting its reliability.
• The agent is interpretable, effectively selects tools, and is deployed in a production platform with over 1600 users.

Conclusions:

• Med.ai ASK dynamically integrates information retrieval tools for accurate and factual biomedical answers.
• The agentic design provides robust interpretative capabilities crucial for the biomedical domain.
• This system represents a significant advancement in AI-assisted scientific research.