💬 NLP

Quantifying Retriever-Generator Alignment in RAG with Local Explanations

This paper introduces RAG-E, an explainability framework that quantifies the often-misaligned interaction between retrievers and generators in Retrieval-Augmented Generation systems using novel attribution methods and the Weighted Attribution-Relevance Gap (WARG) metric to enable more transparent and reliable deployments.

Original authors: Korbinian Randl, Guido Rocchietti, Aron Henriksson, Ziawasch Abedjan, Tony Lindgren, John Pavlopoulos

Published 2026-07-08

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Korbinian Randl, Guido Rocchietti, Aron Henriksson, Ziawasch Abedjan, Tony Lindgren, John Pavlopoulos

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 have a very smart but slightly scatterbrained research assistant (the Generator) who is trying to answer your questions. To help them, you hire a librarian (the Retriever) whose job is to find the best books from a massive library and hand them to the assistant.

In a perfect world, the librarian hands over the most relevant books, and the assistant reads them carefully to write the perfect answer. But in reality, things often go wrong: the librarian might hand over the wrong books, or the assistant might ignore the good ones and make up an answer based on a random page they found earlier.

This paper introduces a new tool called RAG-E (Retrieval-Augmented Generation Explainer) to act like a "black box" inspector. It shines a light on exactly what the librarian and the assistant are thinking, revealing how well they are actually working together.

Here is a breakdown of their findings using simple analogies:

1. The Problem: The "Silent Disconnect"

The authors found that the librarian and the assistant often speak different languages.

The Librarian (Retriever): This part of the system is like a keyword hunter. If you ask, "Where was Marie Curie born?", the librarian scans for the words "Marie Curie" and "born." It grabs the documents that contain those exact words, even if the context isn't perfect.
The Assistant (Generator): This is the AI that writes the answer. The authors found that the assistant treats documents in a binary way: it either thinks a document is "Super Important" (and uses it heavily) or "Total Noise" (and ignores it completely). It doesn't really care about the librarian's ranking.

The Result: The assistant often ignores the librarian's #1 pick (wasting the librarian's effort) or gets distracted by a document the librarian thought was irrelevant (Noise Distraction). In many cases, this mismatch happens in over 70% of the top searches.

2. The Solution: RAG-E (The Inspector)

To fix this, the team built RAG-E, a framework that acts like a magnifying glass for the AI's brain.

For the Librarian: They used a technique called Integrated Gradients. Imagine slowly turning down the volume on every word in a document to see which words were actually "loud" enough to make the librarian pick that book. They discovered that using a specific "unknown" token (like a placeholder for missing info) as a baseline gave the clearest picture of what the librarian was looking at.
For the Assistant: They used a method called PMCSHAP (a fancy, stabilized version of Shapley Values). Imagine playing a game where you remove one book at a time from the assistant's pile to see how much the final answer changes. If removing a book changes the answer drastically, that book was crucial. They found that doing this repeatedly (Monte Carlo style) gave a much more stable and accurate reading than previous methods.

3. The New Scorecard: WARG

The team created a new metric called WARG (Weighted Alignment between Retriever and Generator).

The Analogy: Imagine the librarian hands you a stack of 10 books, ranked from "Best" to "Worst." The assistant then writes an answer. WARG asks: "Did the assistant actually use the books the librarian thought were best?"
Why it's better: Old ways of measuring this were like trying to match two lists of numbers using a straight line (Pearson correlation). But since the assistant treats documents as either "Used" or "Ignored," a straight line doesn't work. WARG is like a specialized ruler that only measures the distance between the "Best" books and the "Ignored" books, giving a much clearer score of how well the team is aligned.

4. What They Found

Misalignment is Common: The librarian and assistant frequently disagree. The assistant often ignores the top-ranked documents or relies on lower-ranked ones.
Keyword Hunting: The librarian is still heavily reliant on matching exact words (nouns and names) rather than understanding the deep meaning of the sentence.
The Score Matters: When the WARG score is high (meaning the librarian and assistant are in sync), the final answer is much more likely to be correct. If they are out of sync, the answer is often wrong.

Summary

This paper doesn't just say "AI is opaque." It builds a tool (RAG-E) to measure exactly how the AI's search and writing parts are failing to agree. They found that these two parts often work at cross-purposes, and they created a new score (WARG) to detect this misalignment, which can help engineers build more reliable AI systems that actually use the information they are given.

Technical Summary: Quantifying Retriever-Generator Alignment in RAG with Local Explanations

Problem Statement

Retrieval-Augmented Generation (RAG) systems combine dense retrievers (RET) and Large Language Models (GEN) to ground outputs in external documents. While this architecture reduces the opacity of LLMs by providing source material, the interaction between the retriever and generator remains opaque. Current RAG systems often exhibit a lack of transparency regarding how information flows between components. Specifically, existing explainability methods typically analyze RET or GEN in isolation, failing to quantify the alignment between them. This creates challenges for deployment in high-stakes domains where understanding whether a generator actually utilizes the most relevant retrieved documents is critical. The authors identify a significant "misalignment" where generators frequently ignore top-ranked documents (Wasted Retrieval) or rely on documents ranked as less relevant (Noise Distraction).

Methodology: The RAG-E Framework

The authors propose RAG-E, an end-to-end explainability framework designed to audit the interaction between RET and GEN components. The framework employs distinct attribution methods tailored to the specific characteristics of each component:

Retriever Analysis (Integrated Gradients):
- The framework applies Integrated Gradients (IG) to dense retrievers.
- A key design choice involves the baseline embedding. Through empirical analysis, the authors found that replacing non-special token embeddings with the model's [unk] token significantly outperforms standard baselines like [pad] or [mask] in terms of faithfulness (measured by Area Inside the Perturbation Curves, AIPC).
- IG is chosen for RETs because they are fully differentiable, allowing for intrinsic, gradient-based explanations that do not scale poorly with input token count.
Generator Analysis (Stabilized Shapley Values):
- For autoregressive generators, the framework uses Shapley Values (SV) to attribute generated tokens to specific context documents.
- To address the computational intractability of exact SVs ( $O(2^{|D|})$ ) and the instability of standard Kernel-SHAP (KSHAP) approximations, the authors introduce PMCSHAP (Paired Monte Carlo Stabilized SHAP).
- PMCSHAP combines Monte Carlo sampling with complementary sampling (pairing input-output sets) to achieve more accurate and reproducible approximations of Shapley Values than native KSHAP, while maintaining manageable computational costs.
- The framework employs Constrained Token Generation to ensure that when evaluating perturbed document subsets ( $D'$ ), the generator's output remains comparable to the original output by constraining the generation sequence based on the original prompt and previous tokens.
Alignment Metric: WARG:
- To answer the research question of how to quantify RET-GEN agreement, the authors propose the Weighted Attribution-Relevance Gap (WARG).
- WARG measures the normalized difference in mean RET importance scores between documents that the generator deems "relevant" (high attribution) and those it deems "irrelevant."
- It utilizes a threshold ( $\tau = 1.5$ ) to identify generation-relevant documents, normalizing attributions and assigning binary weights.
- The metric is clipped to $[-1, 1]$ , where higher values indicate better alignment (the generator focuses on documents the retriever ranked highly).

Key Results

Experiments were conducted on three open-domain datasets (PopQA, QAMPARI, TREC CAST) using two dense retrievers (DRAGON, Arctic Embed 2) and three generator models (Llama 3.1, Qwen 2.5, Gemma 3).

Systematic Misalignment: The study reveals substantial misalignment between RET and GEN. Depending on the model and setting, generators often ignore top-ranked documents. For the top-3 documents, misalignment (Wasted Retrieval or Noise Distraction) affects over 70% of queries in some configurations.
Retrieval Behavior: Quantitative analysis shows that even dense retrievers rely heavily on keyword matching, particularly nouns and proper nouns, rather than deep semantic understanding.
Generator Behavior: Generators exhibit a near-binary attribution pattern: relevant documents receive consistently high scores, while irrelevant documents cluster near zero.
Metric Superiority: WARG outperforms standard Pearson and Spearman correlations in capturing RET-GEN alignment. Unlike correlations, WARG is robust to the near-binary structure of generator importance and the noise in tail distributions.
Performance Correlation: High WARG scores correlate with downstream task performance. Generated answers containing ground-truth labels are more than twice as likely to achieve a high WARG score ( $>0.5$ ) compared to those that do not.

Significance and Claims

The paper claims that RAG-E and WARG provide a practical framework for auditing RAG systems, addressing a gap where previous work focused on components in isolation. The significance of this work lies in:

Transparency: Making the interaction between retrieval and generation explicit and measurable, which is crucial for high-stakes domains like healthcare and law.
Diagnostic Capability: Identifying specific failure modes (Wasted Retrieval, Noise Distraction) that standard metrics might miss.
Optimization Potential: The authors suggest that WARG can serve as an indicator of RAG performance and potentially guide system optimization, such as reducing computational costs by retrieving fewer documents if alignment is known to be poor.
Methodological Contribution: The introduction of PMCSHAP for stabilizing Shapley Value approximations in autoregressive models and the validation of the [unk] token as a superior baseline for Integrated Gradients in retrieval tasks.

The authors maintain modesty regarding their claims, noting that while high WARG correlates with correct answers, it is not a mathematical guarantee of factual correctness. They also acknowledge limitations, including the computational cost of PMCSHAP for very large pipelines and the current focus on standard RAG architectures without rerankers or hybrid retrieval systems.