Evaluation

When applying evaluation, consider the following approaches:

1. Offline Evaluation
   Suitable for prompts requiring a known reference answer (e.g., correctness checks).
2. Online Evaluation
   Used for prompts without a strict reference, letting you assess the system in real-time scenarios.
3. Pairwise Evaluation
   Compares answers from different RAG chains or configurations based on user-defined criteria such as format or style.

Pointwise vs Pairwise Evaluation

Source: Arxiv - Large Language Models for Information Retrieval

Pointwise methods measure relevance between a query and a single document. Subcategories include relevance generation and query generation, both effective in zero-shot document reranking.

4 Different application-specific techniques

1. Agents

• Definition & Role
  • Agents are LLM-driven systems that can take actions and use external tools based on user requests or conversation context.
  • They often employ a planning-and-execution loop, determining when to consult APIs, perform calculations, or retrieve specific documents.
• Evaluation Challenges
  • Multi-Step Reasoning: Agents must plan tasks effectively; improper step-by-step reasoning can lead to incorrect or suboptimal results.
  • Tool Integration: An agent’s effectiveness depends on how accurately it invokes external tools.
  • Error Compounding: A minor mistake early in the chain of thought can derail subsequent steps, making it tricky to evaluate correctness at each stage.
• Evaluation Methods
  • Step-by-Step Trace Analysis: Examining the chain of thought and verifying each action (e.g., which APIs are called, which documents are retrieved) reveals where logic might fail.
  • Scenario Testing: Setting up test scenarios (like booking a flight or summarizing search results) to see if the agent follows correct reasoning steps and returns the right final answer.
  • Human or Heuristic Checks: Human judges or rule-based checks ensure the agent’s output is consistent with each intermediate action (e.g., confirming the correctness of tool usage).

2. Retrieval-Augmented Generation (RAG)

• Core Idea
  • RAG systems retrieve relevant documents from a large corpus, then feed those documents to an LLM to produce a context-grounded answer.
  • The goal is to reduce hallucination by anchoring responses in specific references.
• Evaluation Focus
  • Retriever Performance: Ensuring the system fetches the right context. Common metrics include relevance, recall, precision, and Mean Reciprocal Rank (MRR).
  • Faithfulness & Accuracy: Checking if the generated output truly reflects the retrieved info instead of introducing fabricated content.
  • End-to-End Quality: Although retriever and generator can be evaluated separately, it’s often best to assess the final user-facing answer for correctness and completeness.
• Methods & Best Practices
  • Reference Answers: For fact-based queries, compare the system’s output to a known gold-standard.
  • Document Relevance Checks: Evaluate how well the retrieved documents match the original question.
  • LLM-as-Judge: Employ a secondary model to grade correctness, groundedness, or relevance of the final answer.

3. Summarization

• Purpose
  • Summaries distill lengthy text—articles, reports, transcripts—into a concise form.
  • The challenge is ensuring the summary is both complete (capturing essential points) and accurate (no fabricated details).
• Common Pitfalls
  • Omission of Key Details: Summaries can accidentally skip crucial information.
  • Hallucinations: The model might invent facts not in the source text, undermining trust.
  • Length Constraints: Some tasks require extremely short summaries, which risk omitting context or nuance.
• Evaluation Strategies
  • Reference-Based Metrics: BLEU, ROUGE, and BERTScore measure overlap with a human-written summary. These provide a baseline but may miss subjective quality issues.
  • Expert Review: Domain experts check fidelity, comprehensiveness, and clarity.
  • Quality Dimensions: Focus on coverage (did it include main points?) and faithfulness (did it avoid factual errors or additions?).

4. Classification & Tagging

• Application Scope
  • Encompasses tasks like labeling text with categories (e.g., sentiment analysis) or assigning tags (e.g., topics, entities, or user-defined attributes).
  • May also involve multi-label classification, where each item can belong to multiple categories.
• Key Considerations
  • Label Consistency: The system should maintain consistent use of labels across examples.
  • Granularity: Labels can be too broad or too narrow, depending on user needs.
  • Edge Cases: Overlapping categories or ambiguous content can lead to confusion and mislabeling.
• Evaluation Metrics
  • Precision, Recall, F1-score: Traditional classification metrics help gauge the system’s correctness.
  • Confusion Matrix Analysis: Identifies which categories are commonly misclassified.
  • Human Validation: Humans can review borderline cases for more nuanced interpretation, especially in subjective categories like sentiment or style.

What is reference-free?

When the documentation says “reference-free,” it means you don’t have a single canonical or gold-standard answer to compare against. In other words, there isn’t a labeled dataset that says “The correct answer must be X.” Instead, you might rely on:

• Human or LLM-based judgment to rate the quality of answers (e.g., is the response helpful, coherent, or relevant?).
• Task- or domain-specific heuristics (e.g., checking if code compiles, or if an explanation covers certain keywords).

For instance, if you’re building a creative writing LLM or an open-ended chatbot, you often don’t have a “right answer” for each query—thus, you’re doing “reference-free” evaluation, focusing on subjective measures like clarity, helpfulness, or style rather than comparing to a single correct solution.

TO-DO jotted

In the documentation, “Offline evaluation” is described as a strategy for assessing prompts or outputs in a non-live environment, typically using a labeled dataset or reference answers. While RAG systems often use offline evaluation for checking retrieval accuracy or final answer correctness, offline evaluation can be applied to any LLM application — not just RAG. For example:

• Agentic LLM applications: You might have a multi-step reasoning agent that works with external tools. You could still perform offline tests on a set of known questions and reference answers or states to verify correctness.
• Translation or Summarization tasks: You can gather reference texts or ground-truth summaries and compare model outputs using automated metrics (ROUGE, BLEU, etc.) offline.
• Open-ended conversation: Even if there’s no single correct answer, you can still do a reference-free offline evaluation by rating quality or style using a known dataset of “good” conversation examples.

So, offline evaluation is not limited to only the RAG portion. It can be used to evaluate any part of an LLM-driven system where you want to test outputs or behaviors against a dataset or reference standard — or simply in a controlled setting outside live user interactions.