Exploring the Integration Strategies of Retriever and Large Language Models

Ye Liu, Semih Yavuz, Rui Meng, Meghana Moorthy, Shafiq Joty, Caiming Xiong, Yingbo Zhou

Abstract

The integration of retrieved passages and large language models (LLMs), such as ChatGPTs, has significantly contributed to improving open-domain question answering. However, there is still a lack of exploration regarding the optimal approach for incorporating retrieved passages into the answer generation process. This paper aims to fill this gap by investigating different methods of combining retrieved passages with LLMs to enhance answer generation. We begin by examining the limitations of a commonly-used concatenation approach. Surprisingly, this approach often results in generating "unknown" outputs, even when the correct document is among the top-k retrieved passages. To address this issue, we explore four alternative strategies for integrating the retrieved passages with the LLMs. These strategies include two single-round methods that utilize chain-of-thought reasoning and two multi-round strategies that incorporate feedback loops. Through comprehensive analyses and experiments, we provide insightful observations on how to effectively leverage retrieved passages to enhance the answer generation capability of LLMs.

Introduction

Retrieval-Augmented Generation (RAG) has become the standard for grounding Large Language Models (LLMs) in external facts. Most implementations rely on a "stuff-the-prompt" approach: simply concatenating the top-$k$ retrieved documents. However, this approach is fundamentally flawed and often leads to two major failure modes:

  1. The "Unknown" Trap: Even if the correct answer is present in the text, the model may become overwhelmed by irrelevant "distractor" passages and respond with "unknown." This happens in roughly 20% of cases with simple concatenation.
  2. The Erroneous Majority: An alternative is "Post-Fusion," where the model looks at each document individually to generate an answer pool and then uses a majority vote. While this reduces "unknown" responses, it often leads to the model picking a popular but incorrect answer if the retriever brings back several similar-looking but wrong documents.
Comparison of Concatenation and Post-Fusion architectures and their respective "Unknown" response rates.

Integration Strategies

To bridge the gap between simple concatenation and robust answer generation, we investigated how to better handle model uncertainty and filter out retrieval noise through two categories of improved integration strategies.

Single-Round Refinements

Instead of just dumping text, these methods use Chain-of-Thought (CoT) prompting to guide the model:

  • Pruning Prompt: The model is instructed to explicitly identify and discard irrelevant passages before attempting to answer.
  • Summary Prompt: The model first summarizes the relevant information from all passages and then uses that synthesis to generate the final answer.

Multi-Round Strategies

These methods use the model's initial output as feedback to trigger a second, more intensive processing step:

  • Post-Fusion as the Fallback (Concat + PF): The model tries the efficient Concatenation method first. If it returns "unknown," the system "falls back" to the more robust Post-Fusion method (processing documents individually) to find the answer.
  • Concatenation as the Distiller (PF + Concat): First, the model uses Post-Fusion to generate a pool of candidate answers and identify which passages were actually useful. Then, it performs a final Concatenation step using only those filtered passages to "distill" the final response.
Diagrams of the proposed multi-round strategies: Post-Fusion as the Fallback and Concatenation as the Distiller.

Key Results and Insights

The multi-round approaches consistently outperformed the baselines across three widely recognized open-domain question-answering datasets: Natural Questions (NQ), TriviaQA, and SQuAD-Open.

  • Accuracy Boost: The Concat + PF strategy improved Exact Match (EM) scores by over 10% on average across three LLMs (Llama-2-7b, GPT-3.5, and GPT-4) compared to standard concatenation.
  • Efficiency: While multi-round methods use more tokens, the Concat + PF method is surprisingly efficient. Because it only triggers the second round when the first round fails, it only adds about 90 tokens on average compared to the baseline.
  • "Unknown" is a Signal: The model saying "I don't know" is often a sign of confusion caused by context length or distractors, not a lack of information. Treating this as a trigger for a fallback strategy is highly effective.
  • The Finder vs. Distiller Roles: Post-fusion is excellent at extracting answers from individual snippets (a "finder"), while concatenation is better at weighing multiple pieces of evidence against each other (a "distiller")—provided the noise has been filtered out first.

Conclusion

Effective RAG requires moving beyond simple concatenation to handle the inherent noise in retrieval. By treating "unknown" responses as a signal for model uncertainty and utilizing multi-round fallback or distillation strategies, we can significantly enhance the answer generation capability of LLMs. Future work will explore how these strategies scale as the number of retrieved passages grows and how they perform on non-textual data formats.

Summary updated

Abstract

The integration of retrieved passages and large language models (LLMs), such as ChatGPTs, has significantly contributed to improving open-domain question answering. However, there is still a lack of exploration regarding the optimal approach for incorporating retrieved passages into the answer generation process. This paper aims to fill this gap by investigating different methods of combining retrieved passages with LLMs to enhance answer generation. We begin by examining the limitations of a commonly-used concatenation approach. Surprisingly, this approach often results in generating "unknown" outputs, even when the correct document is among the top-k retrieved passages. To address this issue, we explore four alternative strategies for integrating the retrieved passages with the LLMs. These strategies include two single-round methods that utilize chain-of-thought reasoning and two multi-round strategies that incorporate feedback loops. Through comprehensive analyses and experiments, we provide insightful observations on how to effectively leverage retrieved passages to enhance the answer generation capability of LLMs.
BibTeX
@misc{liu2023exploring,
  author = {Ye Liu and Semih Yavuz and Rui Meng and Meghana Moorthy and Shafiq Joty and Caiming Xiong and Yingbo Zhou},
  title = {{Exploring the Integration Strategies of Retriever and Large Language Models}},
  year = {2023},
  eprint = {2308.12574},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2308.12574},
}
Figure