RAG-MedDx: A Preliminary Offline Prototype of a Citation-Traced Retrieval-Augmented Framework for Intelligent Medical Diagnosis Support

In the modern world of medicine, computers have become powerful assistants, capable of reading vast libraries of medical text and speaking in fluent, human-like sentences. These systems, known as large language models, work by predicting the next word in a sentence based on patterns they have learned from millions of documents. While they are impressive at sounding confident and coherent, they have a significant flaw: they do not always know the truth. Because they rely on memory rather than a live check of facts, they can sometimes invent details that sound plausible but are completely wrong. In a high-stakes field like healthcare, where a wrong guess about a disease or a drug interaction could endanger a patient, this tendency to make things up is a serious problem. Furthermore, when these computer systems give an answer, they often act as black boxes, offering no way to see which book or study they used to reach that conclusion. Doctors and patients need to know not just what the answer is, but where it came from, so they can verify it.

To address these dangers, a researcher named Akanksha Gaikwad has built a preliminary prototype of a new kind of medical support system. This system, called RAG-MedDx, is designed to force the computer to look up the answer in a specific set of trusted books before it speaks, and then to show exactly which book it used. The goal is to create a diagnostic tool that does not just guess, but traces every single statement back to a real source. The researcher tested this idea by creating a simplified, offline version of the system that runs entirely on a local computer without connecting to the internet. The results show that the system can successfully find relevant medical information, link it to a specific question, and admit when it does not know the answer, all while avoiding the need for a cloud-based computer to do the thinking.

The core idea behind this work is to change how a computer generates medical advice. Instead of letting the computer rely solely on its internal training, the new system acts like a librarian who must check the shelves before answering a question. When a user types in a description of symptoms, the system first searches a small, local collection of six medical textbook entries. It looks for keywords that match the user's query. If it finds a match, it pulls that information out and uses it to build an answer. Crucially, every sentence in the final answer comes with a small tag that points directly to the specific textbook entry it came from. This creates a clear path from the diagnosis back to the evidence, allowing a human to verify the source immediately. If the system searches its collection and finds nothing that matches the symptoms, it does not try to guess. Instead, it simply tells the user that no matching information was found, preventing the system from making up a diagnosis.

In the tests conducted for this paper, the prototype demonstrated that this approach works in a controlled environment. When a user entered a query about chest pain radiating to the left arm, the system quickly searched its local database. It found two relevant sources and generated an answer in 0.77 seconds. The response included inline citations that linked each part of the advice to the specific textbook entry it was based on. In a second test, the user entered a misspelled word that did not match any of the medical terms in the database. In this case, the system correctly reported that no sources were found and did not generate a fake answer. The entire process took 0.49 seconds. These tests were run on a single device without any connection to external servers, proving that the system can function offline and keep patient data private.

It is important to understand the scope of what this paper has achieved. The system described is a prototype, a first step in a larger project. The version tested in this study did not use a real, advanced artificial intelligence to write the answers; instead, it used a simplified set of rules to simulate how an AI might behave. Because of this, the paper does not claim that the system is ready to diagnose real patients or that it is better than current medical tools. The researchers have not yet measured whether this method reduces errors or improves the speed of diagnosis compared to standard computer models. The work presented here is a demonstration of feasibility. It shows that it is possible to build a system that retrieves information, attaches citations, and refuses to answer when it lacks evidence, all within a single, inspectable pipeline.

The researcher also built a separate computer framework to test the full system in the future. This framework is designed to compare the new method against a standard computer model that does not look up information. The plan is to run both systems on a large set of medical questions and measure how often they are right, how often they make things up, and how long they take to answer. However, these tests have not been run yet. The current paper focuses on the design of the system and the successful demonstration of its basic functions. The researcher acknowledges that the current prototype uses simple keyword matching, which can miss answers if a word is spelled incorrectly, and that the rule-based answer generation is a stand-in for a more complex intelligence.

Despite these limitations, the prototype offers a clear vision for how medical artificial intelligence could be made safer and more trustworthy. By requiring the computer to show its work and admit when it does not know, the system addresses two major concerns: the risk of false information and the lack of transparency. The ability to run entirely offline suggests that such a system could be used in hospitals without sending sensitive patient data to the cloud. While the work is preliminary and the final results on accuracy and reliability are still to come, the prototype successfully proves that a citation-traced, no-match-aware retrieval system can be built and tested. It lays the groundwork for a future where medical advice from computers is not just fluent, but also verifiable and honest.