Vulnerabilities of a Medical and Scientific based RAG System

When dealing with content like Medical or scientific, things become more complex. Nowadays, Software engineers are increasingly adding semantic search capabilities to applications using RAG. A RAG system involves finding documents that semantically match a query and then passing the documents to a large language model (LLM) to extract the right answer using an LLM. RAG systems aim to reduce the problem of LLM hallucination and provide appropriate links to sources or references for the generated responses and remove the need for annotating documents with meta-data. However, RAG systems suffer from limitations inherited from the quality of retrieved information and from LLMs limited input context window to name a few. In this paper, we present an experience report on the limitations of RAG systems from multiple case studies from separate domains: medical and scientific. We share the lessons learned and show how to prevent those limitations to consider when designing a RAG system. We show that an enhancement of up to 19% could be added on RAG systems when implementing those suggestions.