AI-POWERED HEALTHCARE ASSISTANT: A RAG-BASED CHATBOT WITH HUGGING FACE TRANSFORMERS AND LANGCHAIN

Abstract

A significant issue in healthcare applications is the tendency of traditional AI language models to deliver convincing but factually wrong or hallucinated responses. This shortcoming has been brought to light by the growing demand for reliablemedical information. False assertions about facts and an absence of citations for reliable sources are consequences of using just parametric data in conventional methodslike GPT-3.5 and other fine-tuned LLMs. In complicated medical domains in particular, current implementations of retrieval-augmented generation (RAG) systems suffer from an imbalance between retrieval accuracy, semantic consistency, and response appropriateness, despite the fact that RAG systems were proposed as a solution to these problems. This research built a RAG-based medical AI assistant to help with these challenges. In order to generate responses quickly, it employs FAISS for vector retrieval and Mistral-7B-Instruct-v0.3. The AI has been programmed to only give responses that are backed by credible medical sources, asstated in The Gale Encyclopedia of Medicine. The quantitative measures that were utilized for the evaluation of our system were BERTScore (Precision: 0.8334, Recall: 0.8119, F1-Score: 0.8225), Answer Relevancy (0.9221), and Faithfulness (1.0). Its performance was much better than that of general-purpose models such as GPT-3.5-Turbo (Faithfulness: 0.89) and LLaMA-2-70B-Chat (Faithfulness: 0.95). While maintaining therapeutic relevance and high semantic consistency, our RAG framework successfully reduces hallucinations, according to the results. The fact that problems still arise when dealing with exceedingly complex or ambiguous inquiries further highlights the necessity for improved reasoning approaches. This paper highlights the possibilities of RAG systems that are customized for particular healthcare fields. Provides a trustworthy means of accessing accurate medical records and opens the door to developments in dynamic 2 knowledge integration and multi-hop retrieval. The results support the use of specialized AI helpers in fields like healthcare and academics where precision, openness, and dependability are critical.