Groundedness and Relevance

Module: Optimize GenAI Systems

Section: Quality Optimization

Lesson: Groundedness and Relevance

Introduction: The Foundation of Trust in GenAI

When we build systems powered by Large Language Models (LLMs), the excitement often centers on the model's ability to generate fluent, human-like text. However, in professional and technical environments, fluency is secondary to accuracy. If a system generates a beautifully written response that contains factual errors or wanders off-topic, it fails to provide value and can actively harm the user's workflow. This is where the concepts of Groundedness and Relevance become the cornerstones of quality optimization.

Groundedness refers to the extent to which a model’s output is supported by the provided source material or external facts. A grounded system does not "hallucinate" or invent information; it acts as a faithful synthesizer of the data it has been given. When we talk about groundedness, we are effectively measuring the "truthfulness" of the model relative to a specific context. If your RAG (Retrieval-Augmented Generation) system pulls a document about company benefits, the model should only reference the specific numbers and policies contained within that document, rather than relying on its internal, potentially outdated training data.

Relevance, on the other hand, measures how well the generated response addresses the user's specific query. Even if a response is perfectly grounded, it is useless if it fails to answer what the user asked. A system might retrieve the correct documents and generate facts that are true, but if it spends three paragraphs explaining the history of a topic when the user just asked for a summary of a specific contract clause, the relevance score is low. Together, groundedness and relevance ensure that your AI system is both honest and helpful.

Understanding Groundedness: The Anchor of Truth

Groundedness is essentially about containment. In a well-grounded system, the model’s output is logically derived from a set of provided "ground truth" documents. When a model deviates from these documents to fill in gaps with its own general knowledge, we encounter a loss of groundedness. This is the primary driver of hallucinations in RAG architectures.

The Three Pillars of Groundedness

1. Source Attribution: Does the model explicitly or implicitly rely on the provided context? If the answer is "no," the model is likely hallucinating.
2. Factuality: Are the claims made by the model verifiable against the source text? Even if the model cites the right document, if it misinterprets a statistic or a policy, it lacks groundedness.
3. Constraint Adherence: Does the model follow the instructions regarding the source material? For example, if you explicitly tell the model "only use the provided text," the model must ignore its internal knowledge base even when that knowledge might be factually correct in the real world.

Callout: Groundedness vs. Accuracy While these terms are often used interchangeably, they have a subtle distinction. Accuracy is a broader concept that asks, "Is this information true in the real world?" Groundedness is more specific: "Is this information supported by the provided context?" You can have a grounded response that is inaccurate if your source documents are wrong, and you can have an accurate response that is ungrounded if the model pulls facts from its training data that aren't in your source material.

Practical Example: Identifying Ungroundedness

Imagine a customer service bot that handles internal HR queries.

• User Query: "What is the policy for remote work in the London office?"
• Retrieved Context: "Employees in the London office are eligible for 2 days of remote work per week, provided they have been with the company for at least 6 months."
• Ungrounded Output: "The London office allows 3 days of remote work per week for all employees."

In this case, the model ignored the context (2 days vs 3 days) and the constraint (6 months tenure). This is a failure of groundedness because the response contradicts the provided source of truth.

Understanding Relevance: The Measure of Utility

Relevance is the bridge between the model's output and the user's intent. While groundedness keeps the model honest, relevance keeps the model focused. A common pitfall in GenAI development is focusing so heavily on retrieval precision that we ignore the user's conversational goal.

The Three Dimensions of Relevance

1. Intent Matching: Does the response address the core question asked by the user? If a user asks "How do I reset my password?" and the model explains how to change a username, the intent match is poor.
2. Conciseness: Does the response provide the answer without unnecessary fluff? Irrelevant information—even if true—dilutes the utility of the response.
3. Completeness: Does the response cover all parts of a multi-faceted question? If a user asks two questions, a relevant response must address both, not just the first one.

Note: Relevance is highly dependent on the "Persona" or "System Prompt" you define. If your system is designed to be a brief, direct assistant, a verbose answer is considered irrelevant. If your system is designed to be a conversational coach, a brief answer might be considered irrelevant because it lacks the necessary context or tone.

Technical Implementation: Measuring and Improving Quality

To optimize for groundedness and relevance, we need to move away from subjective evaluation and toward programmatic metrics. Using an LLM-as-a-judge pattern allows us to automate the evaluation of these two qualities.

Step-by-Step: Evaluating with an LLM-as-a-Judge

The "LLM-as-a-judge" approach uses a stronger model (e.g., GPT-4 or Claude 3.5 Sonnet) to evaluate the outputs of your primary model.

Step 1: Define the Evaluation Prompt You need to create a system prompt that instructs the judge to score the output based on specific criteria.

# Evaluation System Prompt
You are an expert evaluator for AI responses. You will be provided with:
1. The User Query
2. The Retrieved Context
3. The AI Response

Task: Evaluate the AI Response based on two metrics on a scale of 1-5:
- Groundedness: Is the response entirely supported by the context?
- Relevance: Does the response directly answer the user's query?

Provide your reasoning before the score.

Step 2: Implement the Evaluation Logic Using a library like langchain or standard Python requests to your model provider, you can run this evaluation pipeline.

def evaluate_response(query, context, response):
    evaluation_prompt = f"""
    Query: {query}
    Context: {context}
    Response: {response}
    
    Evaluate based on Groundedness and Relevance (1-5 scale).
    """
    # Call your LLM API here
    result = call_llm(evaluation_prompt)
    return parse_result(result)

Step 3: Iterative Refinement If your evaluation shows low groundedness, you must modify your Retrieval strategy or your System Prompt. If relevance is low, you likely need to refine your prompt to better define the expected structure of the response.

Best Practices for Groundedness

To ensure your system remains grounded, you must control the input stream and the model's behavior during generation.

1. Strict System Instructions: Use clear directives in your system prompt such as: "Use ONLY the provided context to answer. If the answer is not in the context, state that you do not have enough information."
2. Citation Requirements: Force the model to cite the specific document or paragraph it used to generate a claim. This makes it easier for users to verify the information and discourages the model from inventing details.
3. Context Window Management: Ensure that you are not stuffing too much irrelevant information into the context window. When the context is too large, models tend to lose focus and rely on their pre-trained weights rather than the provided text.
4. Post-Generation Verification: Implement a secondary check where the model reads its own output and compares it against the retrieved context to verify that every sentence is supported.

Tip: If your model struggles with groundedness, try "Chain of Thought" prompting. Ask the model to first extract the relevant facts from the context, and then draft the response based only on those extracted facts. This separates the retrieval/verification step from the generation step.

Best Practices for Relevance

Relevance is improved primarily through prompt engineering and better retrieval.

1. Query Expansion/Rewriting: Users often write ambiguous queries. Use a small model to rewrite the user's query into a more specific, search-optimized format before sending it to the retrieval system.
2. Context Filtering: Before sending retrieved documents to the LLM, use a "re-ranking" step. This ensures that only the most relevant documents are included in the prompt, reducing noise.
3. Defining the Goal: Explicitly state the desired output format. If you want a list, ask for a list. If you want a code snippet, ask for a code snippet. Ambiguity in the request leads to irrelevant, rambling responses.
4. User Feedback Loops: Capture implicit and explicit feedback. If users frequently click "thumbs down" or rewrite their queries immediately after an answer, your relevance is likely suffering.

Common Pitfalls and How to Avoid Them

1. The "Defaulting to Training Data" Trap

Models are trained to be helpful, which means they are biased toward answering questions even when they don't know the answer.

• The Fix: Use "negative constraints." Explicitly tell the model: "Do not use outside knowledge. If the answer is not present in the provided text, respond with 'I cannot answer this based on the available information.'"

2. The "Context Overload" Problem

When you provide 20 pages of context, the model may experience "lost in the middle" phenomena, where it forgets the information in the middle of the context window.

• The Fix: Implement a more precise retrieval system, such as a Vector Database with semantic search, to ensure you are only passing the most relevant 3-4 paragraphs to the model.

3. Ignoring the "System Persona"

Developers often change the system prompt but forget how it impacts the model's tone and relevance.

• The Fix: Maintain a "Golden Dataset" of questions and ideal answers. Every time you change your system prompt, run your test suite against this dataset to ensure you haven't accidentally broken the relevance or groundedness of your responses.

Quick Reference: Quality Optimization Matrix

Warning: Be careful with high "temperature" settings. While high temperature (e.g., 0.7+) can make a model feel more "creative," it is detrimental to groundedness. For RAG systems where accuracy is paramount, always set your temperature to 0 or as close to 0 as possible.

Advanced Techniques for Quality Assurance

As your system scales, manual review becomes impossible. You should transition to a tiered evaluation strategy that balances cost, speed, and accuracy.

Tier 1: Deterministic Checks (Code-based)

Before you even call an LLM, use standard code to filter output.

• Regex Checks: If the output must contain a specific format (e.g., a ticket number), use regex to verify it exists.
• Length Constraints: If the response is over a certain token count, flag it for manual review.

Tier 2: Small Model Evaluation (The "Judge" Model)

Use a smaller, faster model (e.g., GPT-4o-mini or a specialized local model like Llama 3 8B) to perform the groundedness check. This is cheaper than using the largest models for every single interaction.

Tier 3: Human-in-the-Loop (HITL)

For high-stakes environments (legal, medical, financial), create a dashboard that surfaces low-confidence responses to human experts. Use the feedback from these humans to fine-tune your prompts or update your retrieval index.

Deep Dive: The Role of Retrieval in Groundedness

We cannot discuss groundedness without addressing the retrieval component. If your retrieval system pulls "garbage" documents (irrelevant or incorrect data), your model will produce "garbage" responses. This is the "Garbage In, Garbage Out" principle applied to GenAI.

To improve retrieval:

• Hybrid Search: Combine keyword-based search (BM25) with semantic vector search. Keyword search is excellent for specific terminology (part numbers, names), while semantic search captures intent.
• Chunking Strategy: If your chunks are too small, the model loses context. If they are too large, the model gets distracted by irrelevant noise. Experiment with "sliding window" chunking or "semantic chunking" that respects document boundaries.
• Metadata Filtering: If your data has clear structures (e.g., by department, by date), use metadata filtering to narrow down the search space before the semantic search runs. This significantly increases the likelihood that the retrieved documents are relevant.

Designing for Failure: Handling "I Don't Know"

A high-quality system is one that knows when it cannot provide an answer. A common mistake is forcing the model to answer even when the context is insufficient. This creates a "forced hallucination" scenario.

To properly handle "I don't know" scenarios:

1. Define the Error Response: Include in your system prompt: "If the context does not contain the answer, reply with a polite 'I'm sorry, I don't have information on that topic in my current knowledge base.'"
2. Confidence Scoring: Some models provide log-probabilities for their tokens. If the average probability of the generated tokens is below a certain threshold, the model is essentially "guessing." You can use this as a trigger to output a fallback message.
3. Fallback Workflow: If the system determines it cannot answer, trigger a secondary workflow. This could be escalating the user to a human agent, suggesting a search in a different knowledge base, or providing a link to a contact form.

Comprehensive Key Takeaways

To ensure your GenAI system maintains high quality, remember these fundamental principles:

• Constraint is Key: Groundedness is not just about the model's intelligence; it is about the constraints you place on it. Always explicitly instruct the model to ignore its internal training data in favor of provided context.
• Measure Everything: You cannot improve what you do not measure. Implement an LLM-as-a-judge pipeline to quantify groundedness and relevance programmatically.
• The Retrieval-Generation Link: A model is only as good as the documents it retrieves. Invest as much time in optimizing your retrieval strategy (chunking, hybrid search, re-ranking) as you do in prompt engineering.
• Iterate with Data: Use a "Golden Dataset" of queries and perfect responses to test your system after every change. This prevents "regression," where a fix for one type of query breaks another.
• Prioritize Conciseness: Relevance is often improved by simply asking the model to be brief. Over-explaining leads to drift, where the model starts hallucinating or losing the thread of the original query.
• Design for "I Don't Know": It is better for a system to admit ignorance than to hallucinate. Building a graceful fallback mechanism is a mark of a professional, production-ready system.
• Temperature Matters: For all factual, grounded tasks, keep your model's temperature at or near zero. Creativity is the enemy of accuracy in a RAG system.

By adhering to these practices, you move your GenAI implementation from a "fun experiment" to a reliable, trustworthy tool that users can depend on for their daily tasks. The focus on groundedness and relevance ensures that the output is not just words on a screen, but actionable, verifiable information.