Retrieval-Augmented Generation (RAG) systems enhance the performance of Large Language Models (LLMs) by incorporating external information fetched from a retriever component. While traditional approaches prioritize retrieving “relevant” documents, our research reveals that these documents can be a double-edged sword. We explore the counterintuitive benefits of integrating noisy, non-relevant documents into the retrieval process. In particular, we conduct an analysis of how diferent types of retrieved documents-relevant, distracting, and random-afect the overall efectiveness of RAG systems. Our findings reveal that the inclusion of random documents, often perceived as noise, can significantly improve LLM accuracy, with gains up to 35%. Conversely, highly scored but non-relevant documents from the retriever negatively impact performance. These insights challenge conventional retrieval strategies and suggest a paradigm shift towards rethinking information retrieval for neural models.
CEUR ceur-ws.org
Large Language Models (LLMs) [
Traditionally, the retrieval component in RAG systems has focused on fetching documents
that are “relevant” to the query [
Our analysis reveals a surprising phenomenon: the inclusion of random documents, often dismissed as noise, can enhance the accuracy of LLM responses. We observe that strategic placement of these random documents within the context can lead to accuracy improvements of up to 35%. In contrast, top-scoring distractor documents, which do not contain the direct answer but are contextually related, can degrade performance by misguiding the model.
These findings suggest a paradigm shift in the design of retrieval strategies for RAG systems. Instead of solely focusing on maximizing relevance, incorporating a balanced mix of document types, including noise, can lead to better overall performance. This counterintuitive approach calls for a re-evaluation of current retrieval methodologies and paves the way for more efective integration of LLMs and retrieval systems.
This study’s contributions can be summarized as follows: • We provide a detailed examination of how diferent types of retrieved documents—relevant, distracting, and random—impact the efectiveness of RAG systems. • We uncover the counterintuitive finding that incorporating random documents, perceived as noise, into the retrieval process can significantly enhance RAG accuracy, with • Our results suggest a need to shift retrieval strategies laying the groundwork for future research to optimize RAG system performance by leveraging both relevance and improvements of up to 35%.
informational noise.
2. RAG
RAG, along with its variations [
The retriever’s role is to find a suficiently small subset of documents the reasoner to answer the query correctly. Among the various retrieval methodologies, the use of a dense retriever has gained prominence due to its efectiveness in handling semantic matches. The dense retriever processes both the query and potential source documents to generate corresponding embeddings ⃗ for the query and ⃗ for each document ∈ . The to allow embedding process can be represented as ⃗= (); ⃗ = ( ) where and
are neural network-based encoders. Once the embeddings are generated, the retrieval process involves computing the similarity, for instance, cosine similarities, between the query embedding and each document embedding. According to these scores, the top-ranked documents are selected for further processing in the generator component.
The second step involves a generator component in charge of synthesizing an answer, typically implemented via an LLM. Generative language models operate by predicting the probability distribution of the next token, given the previous tokens. For a given sequence of words 1, 2, … , , a generative language model aims to maximize the likelihood of this sequence, expressed using the chain rule of probability: ( 1, 2, … , ) = ∏=1 ( | 1, 2, … , −1 ) where ( | 1, 2, … , −1 ) is the conditional probability of the word given the preceding sequence of words 1, 2, … , −1 . In RAG, the generative language model takes a query and the retrieved documents as input and generates a response by sequentially predicting the next token in the sequence. More formally, ( |) ≈ ∏ ∑∈ (|) ( |, ,
1∶−1 ) where (|) (truncated) probability distribution for the top-scoring documents, and ( |, , 1∶−1 ) is a probability distribution parameterized by that generates a current token based on the previously generated tokens, the query, and the retrieved document; this role is filled by the LLM. In the case of dense retrieval, the probability distribution for the top-scoring documents may is the retrieval component that provides a assume a functional form of the kind (|) ∝ exp(⃗⋅ ⃗) .
We perform our analysis on the open-domain version of Natural Questions (NQ-open) [20, 21]
dataset, and we show results for Llama 2 7B-Chat [
In this study, we explored how diferent types of retrieved documents afect RAG systems, focusing on the qualities that a retriever should have to enhance prompt efectiveness for RAG configurations. Our findings challenge the prevailing assumptions about document retrieval. Specifically, we discovered that highly ranked retrieved documents that lack the answer can actually be detrimental to the efectiveness of LLMs. Intriguingly, we found that introducing completely random documents can boost the accuracy of these systems. These results warrant a rethinking of traditional IR systems to better suit emerging NLP systems. In our future work, we plan to investigate whether this behavior is consistent across diferent types of datasets and tasks, involving various models.
Acknowledgments This work is supported by the Spoke “FutureHPC & BigData” of the ICSC – Centro Nazionale di Ricerca in High-Performance Computing, Big Data and Quantum Computing, the Spoke “Human-centered AI” of the M4C2 - Investimento 1.3, Partenariato Esteso PE00000013 - “FAIR Future Artificial Intelligence Research”, SERICS (PE00000014), IR0000013 - SoBigData.it, funded by European Union – NextGenerationEU, the FoReLab project (Departments of Excellence), and the NEREO PRIN project funded by the Italian Ministry of Education and Research Grant no. 2022AEFHAZ. This work was carried out while Florin Cuconasu was enrolled in the Italian National Doctorate on Artificial Intelligence run by the Sapienza University of Rome. reinforced retrieval augmented machine learning, in: R. Basili, D. Lembo, C. Limongelli, A. Orlandini (Eds.), Proceedings of the Discussion Papers - 22nd International Conference of the Italian Association for Artificial Intelligence (AIxIA 2023 DP) co-located with 22nd International Conference of the Italian Association for Artificial Intelligence (AIxIA 2023), Rome, Italy, November 6-9, 2023, volume 3537 of CEUR Workshop Proceedings, CEURWS.org, 2023, pp. 29–37. URL: https://ceur-ws.org/Vol-3537/paper4.pdf. [20] T. Kwiatkowski, J. Palomaki, O. Redfield, M. Collins, A. Parikh, C. Alberti, D. Epstein, I. Polosukhin, J. Devlin, K. Lee, K. Toutanova, L. Jones, M. Kelcey, M.-W. Chang, A. M. Dai, J. Uszkoreit, Q. Le, S. Petrov, Natural questions: A benchmark for question answering research, Transactions of the Association for Computational Linguistics 7 (2019) 453–466. [21] K. Lee, M.-W. Chang, K. Toutanova, Latent retrieval for weakly supervised open domain question answering, in: Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics, Association for Computational Linguistics, Florence, Italy, 2019, pp. 6086–6096. URL: https://www.aclweb.org/anthology/P19-1612. doi:10.18653/ v1/P19- 1612.