Retrieval-Augmented Generation allows to enhance Large Language Models with external knowledge. In response to the recent popularity of generative LLMs, many RAG approaches have been proposed, which involve an intricate number of different configurations such as evaluation datasets, collections, metrics, retrievers, and LLMs. Inconsistent benchmarking poses a major challenge in comparing approaches and understanding the impact of each component in the pipeline. In this work, we study best practices that lay the groundwork for a systematic evaluation of RAG and present BERGEN, an end-to-end library for reproducible research standardizing RAG experiments. In an extensive study focusing on QA, we benchmark different state-of-the-art retrievers, rerankers, and LLMs. Additionally, we analyze existing RAG metrics and datasets.
BERGEN: A Benchmarking Library for Retrieval-Augmented Generation
| David Rau, Hervé Déjean, Nadezhda Chirkova, Thibault Formal, Shuai Wang, Vassilina Nikoulina, Stéphane Clinchant |
| Findings of the Conference on Empirical Methods in Natural Language Processing (EMNLP), Miami, Florida, 12-16 November, 2024 |
Post by author Stéphane Clinchant, November 6, 2024 (LinkedIn)
When we started working on RAG in 2023, we read many research papers and ended up being clueless about what were good practices or simply what could be a good baseline to build on. RAG experimental setups are fragmented, therefore comparing results across papers leads to comparing apples and oranges:
That’s why with David Rau Shuai Wang Hervé Déjean, Nadezhda Chirkova, Thibault Formal, Vassilina Nikoulina we built BERGEN, an open source library to ease the reproducibility of RAG experiments with a set of recommendations for strong baselines in RAG.
The problems with metrics
There are a variety of metrics that papers report to evaluate their models, such as Exact Match (EM), i.e the produced string is exactly the expected answer, Match (M) the expected answer is contained in the generated string , even though sometimes it is now called accuracy by some recent papers. Others prefer F1 or precision. However, it has already been reported (Kamalloo, E. et al), that some of these lexical metrics correlate poorly with human judgements..
We use an LLM-as-a-Judge based on SOLAR-10.7B from Upstage , referred as LLMeval. We then compute correlation with GPT-4 as a judge. LLMeval closely aligns with GPT-4’s evaluation, followed by Match and Recall, making them the most effective non-commercial metrics for (zero-shot) RAG evaluation, among the ones tested.
Which datasets are interesting to benchmark RAG systems?
Now that we selected a metric, we can look at different QA datasets. We examine more than 10 datasets and compare the performance with RAG and without RAG for different LLMs. The results suggest that ASQA, HotpotQA, NQ, TriviaQA, POPQA are the most interesting ones, while some KILT datasets such as ELI5, WoW are not relevant for RAG.
RAG is also about Retrieval, right?
It’s surprising to see how many RAG papers have a poor retrieval setup, using outdated retrievers and no reranker at all. The following graph simply shows the relation between retrieval performance and generation performance to convince you to use a strong retrieval setup (before reinventing reranking with your big LLMs).
Better Baselines
Thanks to our 500+ experiments, we identify a strong baseline for academic RAG datasets (all the results are in the paper)
- top_k documents: 5
- First Stage Retrieval: splade-v3
- Reranker: DeBERTa-v3
- LLM: SOLAR 10.7B from Upstage
Furthermore, we also add multilingual RAG datasets and experiments (cf https://arxiv.org/abs/2407.01463), where we recommend
- Retrieval & Reranker: BGE-m3
- LLM: Command-R from Cohere combined with language-specific prompts
- Collection: multilingual Wiki
Conclusion
While there exists more application oriented libraries for RAG such as LangChain, the situation for academic RAG is more complex and it is hard to compare papers and to know the current state of the art on existing datasets. We hope that BERGEN can help build strong baselines, and better RAG systems!
To make robots autonomous in real-world everyday spaces, they should be able to learn from their interactions within these spaces, how to best execute tasks specified by non-expert users in a safe and reliable way. To do so requires sequential decision-making skills that combine machine learning, adaptive planning and control in uncertain environments as well as solving hard combinatorial optimisation problems. Our research combines expertise in reinforcement learning, computer vision, robotic control, sim2real transfer, large multimodal foundation models and neural combinatorial optimisation to build AI-based architectures and algorithms to improve robot autonomy and robustness when completing everyday complex tasks in constantly changing environments.
The research we conduct on expressive visual representations is applicable to visual search, object detection, image classification and the automatic extraction of 3D human poses and shapes that can be used for human behavior understanding and prediction, human-robot interaction or even avatar animation. We also extract 3D information from images that can be used for intelligent robot navigation, augmented reality and the 3D reconstruction of objects, buildings or even entire cities.
Our work covers the spectrum from unsupervised to supervised approaches, and from very deep architectures to very compact ones. We’re excited about the promise of big data to bring big performance gains to our algorithms but also passionate about the challenge of working in data-scarce and low-power scenarios.
Furthermore, we believe that a modern computer vision system needs to be able to continuously adapt itself to its environment and to improve itself via lifelong learning. Our driving goal is to use our research to deliver embodied intelligence to our users in robotics, autonomous driving, via phone cameras and any other visual means to reach people wherever they may be.
This web site uses cookies for the site search, to display videos and for aggregate site analytics.
Learn more about these cookies in our privacy notice.
Cookie settings
You may choose which kind of cookies you allow when visiting this website. Click on "Save cookie settings" to apply your choice.
This content is currently blocked. To view the content please either 'Accept social media cookies' or 'Accept all cookies'.
For more information on cookies see our privacy notice.