A Practical Guide to RAG Pipeline Evaluation (Part 2: Generation)

How to evaluate different aspects of an answer?

Generation evaluation is all about assessing LLM generated answers. This is different from Retrieval evaluation where the task is to assess the quality of Information Retrieval (which may or may not involve LLMs). Evaluating Generation can help you answer the following questions:

• Do I have to use GPT-4 or would a smaller model work too?
• Should I fine-tune an LLM for my RAG application?
• Which prompts minimize hallucination the most?
• How sensitive are answers to different prompts?
• Is the LLM already good enough if I provide the right contexts, and should I focus on improving Retrieval instead?

It is, however, more challenging to evaluate Generation than to evaluate Retrieval, because what good means is often subjective and context-dependent (contrary to retrieval, where there usually is a right answer to which chunks should be fetched). Nevertheless, there are two common approaches in RAG evaluation, and these can be categorized based on whether they use the ground truth answers or not.

We are going to use the following notation:
Q: Question | C: Retrieved Context(s) | C*: Ground Truth Context(s) | A: Generated Answer | A*: Ground Truth Answer(s) | X ~ Y: Evaluate X w.r.t. Y

The first is to holisticallyjudge the Generated Answer (A) with respect to the ground truth answers (A*):

• Correctness (A~A*): If you have ground truth answers, you can directly compare the two and see how far off the generated answer is from the ground truth. This straightforward measurement holistically captures all the qualities you want for your answer (grounded on context, relevant to question, correct reasoning, ideal style, …).

The second is to judge specific aspects of the Generated Answer (A) with respect to the Question (Q) and Retrieved Contexts (C),WITHOUT ground truth answers (A*):

• Faithfulness (A~C): Is the LLM using the relevant retrieved context to provide an answer? This is important because otherwise, it could be using unvalidated prior knowledge, or even worse, hallucinated information. In the cases where the context is not relevant or complete enough, the LLM should be able to recognize the situation and provide some version of “I don’t know” or “I don’t have all the information to provide an answer”.
• Relevance (A~Q): Does the answer provide a direct response to the question? There could be cases where the answer is just repeating the context, but not answering the question.
• Logic (A~Q, C): Is the LLM able to give the correct answer derived logically from the question and context?
• Style (A~Q or A alone): Is the answer too short, too long, courteous enough, quoting evidence, or summarizing context? What a good style means comes down to user preferences.
• … this is not an exhaustive list as there could be more aspects important to your use case

Now that we know how to evaluate Generation holistically or in specific aspects, let’s take a look in the next section, the specific metrics available to measure them.

What Metrics to use?

We can find a multitude of metrics that measure the various aspect of a generated answer (e.g., correctness, faithfulness, etc). We identified three categories: Deterministic, Semantic and LLM-basedmetrics.

Deterministic metrics like ROUGE, Token Overlap, BLEU simply compares the tokens between the Answer and Reference (C, Q, or A*) and calculates the overlap. The benefit for these metrics is that they can compute almost instantly, but the downside is that they don’t account for semantic meaning.

Semantic metrics leverages small language models purpose-built to perform particular tasks. We find a particular version of DeBERTa model (DeBERTa-v3 NLI) particular useful one in comparing sentence pairs (S1, S2) because it outputs three qualities of the answers (Contradiction: does S1 contradict S2; Entailment: does S1 imply S2; Neutral: if S1 and S2 has no logical relationship). We can use Entailment to assess for if the generated answer (A) implies ground truth answer (A*), and Contradict to see if they have conflicting statements.

LLM-based metrics are the most versatile and can technically evaluate any aspect of the answer. The key is to provide the right prompt and use a capable enough model that is different from the original model used to generate the answer.

How good are these Metrics vs. Human Labels?

As we did in Part 1 of this Guide studying retrieval metrics, we want to understand how practical the generation metrics are, or how aligned they are with human assessment.

Intuitively, we should expect LLM-based metrics to align the best with human assessment because they can capture the reasoning and logic behind the words. But, in the following analysis, we show that deterministic and semantic metrics are not that far off from LLM-based metrics and can also offer effective assessment of answer quality.

To this end, we use the dataset released by the authors of the paper Evaluating Correctness and Faithfulness of Instruction-Following Models for Question Answering. The data includes human-labeled ground truths for Faithfulness (600 examples) and Correctness (1200 examples) evaluation.

Our first step is to understand how the metrics correlate to human labels. We use three standard correlation measures

• Pearson Correlation: It tells us how much two things change together (like height and weight) in a consistent way.
• Spearman’s Rank Correlation: It measures how well the relationship between two things can be described by a consistent trend, using their ranks (like comparing two lists of favorite sports teams).
• Kendall’s Tau: It checks how similar the order or ranking of two sets of things is (like two different top 10 lists of movies).

Here’s the result for Correctness (A~A*):

We can see that LLM-based metrics correlate with human labels the best at assessing correctness, with GPT-4 achieving the highest correlation with human evaluation. Notably, the best Deterministic and Semantic metrics achieve comparable performance to LLM-based Metrics other than GPT 4 Correctness

Here’s the result for Faithfulness (A~C):

Similarly, in Faithfulness evaluation (A~C) we can see that simple deterministic metrics are not that far off from LLM-based metrics, which only work well with complex by_statement_prompt. (by_statement_prompt breaks generated answer (A) into statements and assess if each is attributable to the retrieved contexts (C). binary_prompt just asks for a binary classification of whether the answer (A) is grounded on the context (C). This prompt significantly underperforms when using a less capable model.)

Overall, the results are aligned with our expectations that LLM-based metrics outperforms traditional deterministic and semantic metrics because LLMs excel at understanding nuances in answers and can assess reasoning quality. But not by that much! This insight is important because it shows simpler metrics can be effective too!

In the next section, we show you how you can combine the best qualities of simple metrics (fast & cheap) and LLM-based metrics (better aligned with humans) to predict human labels.

How to use these Metrics in practice?

We concluded from the correlation analysis above that LLM-based generation metricsalign the best with human judgement. But in practice, they are can be costly and slow to run if you want to routinely experiment with different configurations of the RAG pipeline.

In this section, we show the steps to create Ensemble Metrics to run cost-effective evaluation. In particular, we developed a hybrid evaluation pipeline that can reduce the cost of evaluation by 15x compared to pure LLM-based evaluation, without quality downgrade.

The aim of ensembling different metrics to predict the human label is to combine the strengths and balance out the weaknesses of individual metrics, ultimately leading to more accurate, robust, and reliable predictions. Each metric might capture different aspects of the data or be sensitive to different patterns, so when we combine them, we often get a more comprehensive view.

Based on this intuition, we first tested three ensemble models, each combining a subset of metrics: Deterministic Ensemble, Semantic Ensemble and Det_Sem Ensemble

From the Precision/Recall graph below, we see a notable improvement with the Deterministic Semantic Ensemble (Det_Sem Ensemble), balancing precision and recall. Remarkably, Det_Sem Ensemble beats all individual metrics on Precision of correctness classification.

Next, we take it one step further by integrating the powerful GPT-4. This Hybrid Pipeline matches the GPT-4 evaluator’s performance by only running GPT-4 on 7% of datapoints, improving cost by 15x!

Here’s how we construct the hybrid pipeline:

• Step 1: use Deterministic Semantic Ensemble to do a first pass on all the datapoints
• Filter out datapoints where the metric cannot decide with confidence, computed with a tool called “Conformal Prediction” (more below)
• Step 2: use GPT-4 to classify those undecided data points (7% of the total test set data points)

Conformal Prediction is a statistical technique that quantifies the confidence level of a prediction. In this case, we are trying to predict whether the answer is correct (or faithful). With conformal prediction, instead of just saying “yes” (or “no”), the model tells us “the answer is correct with probability at least 90%”. In essence, conformal prediction doesn’t just give you an answer; it tells you how confident you can be in that answer. If the model is uncertain, conformal prediction will tell you it’s “undecided”. For the undecided datapoints, we ask the more powerful GPT-4 to judge its correctness.

Conformal prediction is easy to use (requires only a small number of data points for calibration) and is model agnostic while providing useful coverage guarantees. Note that conformal prediction and also be used on LLM outputs (if they provide log-probabilities). This recent paper Uncertainty Quantification for Black-Box LLMs details some approaches.

Key Takeaways for Generation Evaluation

• Evaluate Retrieval and Generation separately. Generation evaluation helps you make pipeline decisions on LLM and Prompt choices.
• Use Golden Answers for holistic evaluation. Comparing with Golden answers is the best way to capture all qualities you want in answers.
• Assess specific aspects of generated answers. Judge faithfulness, relevance, logic, style, and any other dimensions important to you.
• Choose metrics that correlate highly with human labels. LLM-based metrics win here, but select deterministic and semantic metrics also preform well and should be leveraged.
• Combine different types of metrics to improve quality and lower cost. Use ensembling and multi-step hybrid pipelines to improve the cost, performance, and scalability of evaluation pipelines.

Lastly, here’s the link to the Github repo: continuous-eval if you want to run the metrics mentioned in the article on your data.