Research-Driven Techniques for Getting the Generation Part Right In Retrieval Augmented Generation
This article is co-authored by Evan Jolley New evaluations of RAG systems are published seemingly every day, and almost all of them focus on the retrieval stage of the framework….

This article is co-authored by Evan Jolley

New evaluations of RAG systems are published seemingly every day, and almost all of them focus on the retrieval stage of the framework. Arize published several pieces covering retrieval, including a detailed analysis of the needle in a haystack test.

However, the generation aspect – how a model synthesizes and articulates this retrieved information – may hold equal if not greater significance in practice. Many use cases we see in production are not simply returning a fact from the context, but also require synthesizing the fact into a more complicated response. Whether it’s transforming numerical data into a narrative analysis, converting abstract concepts into tangible examples, or integrating disparate pieces of information, the generation process is essential to the functionality of retrieval-augmented generation systems.

We ran several experiments to evaluate and compare GPT-4, Claude 2.1 and Claude 3.0 Opus’s generation capabilities. This article details our research methodology, results, and model nuances encountered along the way, and why this matters to people building with generative AI.

Takeaways

  • Although initial findings indicate that Claude outperforms GPT-4, subsequent tests reveal that with strategic prompt engineering GPT-4 demonstrated superior performance across a broader range of evaluations. Inherent model behaviors and prompt engineering matter A LOT in RAG systems.
  • Simply adding “Please explain yourself then answer the question” to a prompt template significantly improves (more than 2x) GPT-4’s performance. It’s clear that when an LLM talks answers out, it seems to help in unfolding ideas. It’s possible that by explaining, a model is re-enforcing the right answer in embedding/attention space.
  • The generation step is equally, if not more crucial to RAG system success than retrieval.

The Phases of RAG and Why Generation Is Critical

steps of retrieval augmented generation

That much recent research in RAG systems is directed towards the retrieval stage is understandable given the importance of locating and extracting relevant information from a vast knowledge base. However, it’s important not to overlook the significance of the generation phase, which plays an equally vital role in the overall performance and usability of these systems.

While retrieval is responsible for identifying and retrieving the most pertinent information, it is the generation phase that takes this raw data and transforms it into a coherent, meaningful, and contextually appropriate response. The generative step is tasked with synthesizing the retrieved information, filling in gaps, and presenting it in a manner that is easily understandable and relevant to the user’s query.

In many real-world applications, the value of RAG systems lies not just in their ability to locate a specific fact or piece of information but also in their capacity to integrate and contextualize that information within a broader framework. The generation phase is what enables LLM RAG systems to move beyond simple fact retrieval and deliver truly intelligent and adaptive responses, whether it’s providing a detailed explanation, offering insights based on multiple sources, or generating creative solutions to complex problems.

Test #1: Date Mapping

The initial test we ran involved generating a date string from two randomly retrieved numbers: one representing the month and the other the day. The models were tasked with:

  1. Retrieving Random Number #1
  2. Isolating the last digit and incrementing by 1
  3. Generating a month for our date string from the result
  4. Retrieving Random Number #2
  5. Generating the day for our date string from Random Number 2

For example, random numbers 4827143 and 17 would represent April 17th.

These numbers were placed at varying depths within contexts of varying length. The models initially had quite a difficult time with this task.

gpt-4 turbo vs claude 2.1 date mapping performance

While neither model performed great, Claude 2.1 significantly outperformed GPT-4 in our initial test, almost quadrupling its success rate. It was here that Claude’s verbose nature – providing detailed, explanatory responses – seemed to give it a distinct advantage, resulting in more accurate outcomes compared to GPT-4’s initially concise replies.

Prompted by these unexpected results, we introduced a new variable to the experiment. We instructed GPT-4 to “explain yourself then answer the question,” a prompt that encouraged a more verbose response akin to Claude’s natural output. The impact of this minor adjustment was profound.

testing retrieval with generation

GPT-4’s performance improved dramatically, achieving flawless results in subsequent tests. Claude’s results also improved, just to a lesser extent.

This experiment not only highlights the differences in how language models approach generation tasks but also showcases the potential impact of prompt engineering on their performance. The verbosity that appeared to be Claude’s advantage turned out to be a replicable strategy for GPT-4, suggesting that the way a model processes and presents its reasoning can significantly influence its accuracy in generation tasks. Overall, including the seemingly minute “explain yourself” line to our prompt played a role in improving the models’ performance across all of our experiments.

Further Testing of Retrieval Generation Tasks

RAG generation tests String Concatenation Money Formatting Date Mapping Modulo Arithmetic

We conducted four more tests to assess prevailing models’ ability to synthesize and transform retrieved information into various formats. These tests included:

  • String Concatenation: Combining pieces of text to form coherent strings, testing the models’ basic text manipulation skills.
  • Money Formatting: Formatting numbers as currency, rounding them, and calculating percentage changes to evaluate the models’ precision and ability to handle numerical data.
  • Date Mapping: Converting a numerical representation into a month name and date, requiring a blend of retrieval and contextual understanding.
  • Modulo Arithmetic: Performing complex number operations to test the models’ mathematical generation capabilities.

Each model exhibited strong performance in string concatenation, reaffirming previous understanding that text manipulation is a fundamental strength of large language models.

testing retrieval with generation gpt-4 vs opus claude

Claude 3 and GPT-4 performed almost flawlessly on the money formatting test. Claude 2.1’s performance was generally poorer overall. Accuracy did not vary considerably across token length, but was generally lower when the needle was closer to the beginning of the context window.

Normal haystack test results: opus claude 3 vs GPT-4

Despite stellar results in the generation tests, Claude 3’s accuracy declined in a retrieval-only experiment. Theoretically, simply retrieving numbers should be an easier task than manipulating them as well – making this decrease in performance surprising. The results are strange enough to suggest there may be some mistake in our process and so we are planning more testing. If anything, this counterintuitive dip only further confirms the notion that both retrieval and generation should be tested when developing with RAG.

Analysis

By testing various generation tasks, we observed that while both models excel in menial tasks like string manipulation, their strengths and weaknesses become apparent in more complex scenarios. LLMs are still not great at math! Another key result was that the introduction of the “explain yourself” prompt notably enhanced GPT-4’s performance, underscoring the importance of how models are prompted and how they articulate their reasoning in achieving accurate results.

These findings have broader implications for the evaluation of LLMs. When comparing models like the verbose Claude and the initially less verbose GPT-4, it becomes evident that the evaluation criteria must extend beyond mere correctness. The verbosity of a model’s responses introduces a variable that can significantly influence their perceived performance. This nuance may suggest that future model evaluations should consider the average length of responses as a noted factor, providing a better understanding of a model’s capabilities and ensuring a fairer comparison.

Overall, the generation step of RAG has received less attention than retrieval, but is just as important. Simple prompt engineering and model nuances can lead to massive differences in RAG performance, and when building these systems, rigorous testing and making choices based on specific use cases are necessary.