Same Question, Different Source, Different Answer: Auditing Source-Dependence in Medical Multi-Source RAG
This paper introduces a novel evaluation framework for multi-source Retrieval-Augmented Generation (RAG) systems that shifts the focus from answer correctness to auditing source-dependence, demonstrating through a transplant patient education benchmark that institutional disagreements are far more prevalent than previously estimated and proposing tools like HERO-QA and a structured judge to systematically measure and manage these inter-source relationships.
Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer
Imagine you are a patient who just had a heart transplant. You have a question: "When can I travel internationally again?" You type this into a smart medical chatbot.
In the old way of building these chatbots, the system would find one "perfect" answer and give it to you with total confidence. But this paper argues that in the real world, there is no single perfect answer. Instead, the answer you get depends entirely on which hospital's rulebook the chatbot happened to pick up.
If the bot picks up the rulebook from Hospital A, it might say, "Wait 3 months." If it picks up Hospital B's book, it might say, "Wait 12 months." Both answers are written confidently, both are cited correctly, but they contradict each other. The old way of testing these bots couldn't see this problem because they were only looking for one "gold standard" answer.
This paper introduces a new way to audit these systems, using a transplant patient education system as a test case. Here is how they did it, explained simply:
1. The "Recipe Book" Collection (TransplantQA)
The researchers gathered 102 different patient handbooks from 23 major transplant centers across the U.S. Think of these as 102 different recipe books for the same dish (post-transplant care). Some chefs (hospitals) say "add salt," others say "no salt," and some don't mention salt at all.
They also collected 1,115 real questions that actual patients had asked on forums. They didn't make these up; they are genuine worries from real people.
2. The "Smart Librarian" (HERO-QA)
To test the system, they built a special retrieval engine called HERO-QA. Imagine a librarian who is trying to find the right page in a massive library.
- If the book is short, the librarian reads the whole book to make sure nothing is missed.
- If the book is huge, the librarian uses a smart map to find the specific chapter, then the specific paragraph, and even checks the neighboring paragraphs to get the full context.
This librarian then asks a powerful AI (the "Generator") to write an answer based only on the pages it found. They did this for every single question against every single handbook. This resulted in over 48,000 different answers to the same questions.
3. The "Tough Judge" (Structured Output)
Now comes the most important part. They needed a way to compare these 48,000 answers to see how they differed. They didn't just ask, "Are they the same or different?" They used a specialized AI "Judge" that acts like a strict editor.
Instead of just giving a score, this Judge writes a short report for every pair of answers it compares. It sorts them into five buckets:
- Absent: One book didn't even talk about the topic.
- Consistent: Both books say the exact same thing.
- Complementary: They agree on the main point but add different details (like one says "eat healthy" and the other adds "and exercise").
- Divergent: They give different advice (e.g., "wait 6 weeks" vs. "wait 12 weeks").
- Contradictory: They say opposite things (e.g., "You can do this" vs. "Never do this").
Crucially, the Judge also explains why they differ and how serious the difference is.
4. The Big Discovery
When they ran the numbers, they found something surprising.
- The "Missing" Problem: In about 79% of the cases, one of the handbooks didn't have an answer at all. The old way of testing missed this because it assumed every book had an answer.
- The "Hidden" Disagreement: When they fixed the retrieval system (making the librarian smarter), they found more disagreement, not less.
- The Analogy: Imagine you are looking for a specific word in a dictionary. If you only look at the first page, you might think everyone agrees on the definition. But if you read the whole dictionary, you realize different editions define it differently.
- The paper found that previous estimates understated how often hospitals disagreed. It wasn't that the disagreement was stronger; it was that the old systems were just blind to the fact that many hospitals didn't have an answer at all, hiding the true extent of the differences.
5. Why This Matters Beyond Medicine
The authors say this isn't just about heart transplants. They argue that any system that pulls answers from multiple sources (like legal systems where laws differ by state, or schools where curriculum standards differ by district) has this same blind spot.
If a student asks a chatbot about history, the answer might change depending on whether the bot is reading a textbook from New York or Texas. This paper provides a toolkit to measure that "source-dependence" so we can stop pretending there is always one single right answer.
In summary: The paper built a massive test to show that when AI answers questions based on multiple sources, the answer often depends on which source it picked. They created a new way to measure these differences, proving that we need to stop looking for a single "correct" answer and start auditing how different sources relate to one another.
Drowning in papers in your field?
Get daily digests of the most novel papers matching your research keywords — with technical summaries, in your language.