AI Product Evaluation Design

AI Product Evaluation Design

In the era of LLMs, product development moves from writing static specifications to defining "correctness" through Evals. Since models are stochastic, you cannot "fix a bug" with a single line of code; instead, you must "hill climb" toward better behavior by building robust datasets that measure model performance against your product goals.

The Three-Tier Evaluation Framework

Depending on the complexity of the feature, use one or more of these evaluation methods:

1. Deterministic Evals (Pass/Fail)

Best for extraction, tool-calling, or objective facts.

• Goal: Verify the model extracts the exact right data.
• Example: If the user says "Remind me to eat at 7 PM," the JSON output for time must be 19:00.
• Metric: Accuracy % (Total correct / Total prompts).

2. Human Preference Evals (Side-by-Side)

Best for tone, creativity, and visual design (like the "Canvas" layout).

• Goal: Compare two model versions (e.g., a baseline vs. a new fine-tuned model).
• Process: Present a prompt and two anonymized completions. Ask a human rater: "Which is better for [Specific Goal]?"
• Metric: Win Rate (The percentage of time the new model beats the baseline).

3. Model-Graded Evals (LLM-as-a-Judge)

Best for scaling quality checks without manual labor.

• Goal: Use a high-reasoning model (like o1) to grade the output of a faster, cheaper model.
• Process: Give the "Judge" model the rubric of what a "good" response looks like and ask it to score the "Student" model on a scale of 1-5.

Step-by-Step Process for Designing Evals

1. Create the "Ground Truth" Dataset

Build a spreadsheet with the following columns to define the model's target behavior:

• Input/Prompt: What the user says (include diverse variations).
• Baseline Behavior: How the current model responds.
• Ideal Behavior: A hand-written "Golden Response" showing exactly what you want.
• Rationale: Why the ideal behavior is better (e.g., "It didn't trigger the UI when it should have stayed in chat").

2. Define Decision Boundaries

For agentic features (like Canvas or Task execution), define the "Trigger Boundary":

• Trigger Scenarios: Prompt: "Write a 5-page essay." Result: Model opens the document editor.
• Non-Trigger Scenarios: Prompt: "Who is the President?" Result: Model stays in the standard chat interface.

3. Identify Performance Regressions

When you optimize for one skill (e.g., "Being more concise"), you may accidentally "brain damage" another skill (e.g., "Formatting code correctly").

• Always run your new feature evals alongside a "General Intelligence" eval set to ensure core reasoning hasn't dropped.

Examples

Example 1: Deterministic Eval for a "Tasks" Tool

• Context: An AI assistant that sets reminders.
• Input: "Remind me to call Mom in two hours." (Sent at 10:00 AM).
• Expected Output: { "action": "set_reminder", "content": "Call Mom", "time": "12:00" }.
• Application: Run 100 variations of time-based language ("tonight," "in a bit," "next Tuesday") to ensure the extraction logic holds.

Example 2: Preference Eval for Writing Style

• Context: Improving the "friendly" tone of a document editor.
• Input: "Rewrite this paragraph to be more encouraging."
• Model A: "You did a good job on the report."
• Model B: "This report is a fantastic start! Your analysis of the data is really sharp."
• Evaluation: Human rater chooses Model B because it uses specific positive reinforcement instead of generic praise.

Common Pitfalls

• Measuring the Wrong Baseline: Using a weak model as your baseline makes your new model look better than it actually is. Always test against the "state of the art" (SOTA).
• Neglecting Diversity: Training or testing only on "happy path" prompts. Include edge cases, slang, and non-English inputs to ensure the model doesn't fail in the wild.
• The "Over-Refusal" Trap: Teaching a model to be too safe or helpful can cause it to start refusing valid requests (e.g., the "body paradox" where a model refuses to set an alarm because it "doesn't have a physical body").
• Ignoring Latency: A model that is 5% more accurate but 10x slower is often a net-negative for the user experience. Always include "Time to First Token" as an eval metric.