
Next, once your UI capture is stable and quick checks pass, you need a scoring layer that turns “seems fine” into a number you can track week to week. The goal is not a perfect grade for every response, but a reliable signal that catches prompt, UI, or model drift before users do.
If you only do one thing, pick 2 to 4 metrics that match how your chatbot can fail in real usage. More metrics can add noise and slow runs, so start with what maps to your biggest risk and expand only when you see gaps.
Also, different use cases call for different grades, and the wrong metric can give you false confidence. A support bot that must cite policy needs different checks than a writing assistant that must follow formatting rules.
A practical starter set looks like this:
Faithfulness: checks whether the answer stays grounded in the provided context, not invented details
Relevance: checks whether the response answers the question asked, not a nearby question
Correctness: checks whether the final answer is actually right for tasks with a clear expected output
Formatting: checks structure requirements like JSON shape, bullet lists, headings, or required fields
Toxicity: checks for harmful or unsafe language where it must never appear
Here’s the catch: some metrics work best when the task has clear reference material, and they fail when the task is open-ended.
Correctness works best when you have an expected answer, a rubric, or a known data source
Correctness fails when the user requests brainstorming or creative writing, so favor relevance and formatting there
Faithfulness works best for RAG (retrieval-augmented generation) flows where you pass documents into the prompt
Faithfulness fails when you do not provide source context, because there is nothing to be faithful to
That said, a score is only useful if you decide what “good enough” means and you keep a baseline run. Treat your first stable run as your reference point, then compare future runs to detect regressions caused by prompt edits, UI changes, or a model version swap.
A simple way to start:
Choose a pass threshold per metric (for example, stricter for toxicity and formatting, looser for relevance)
Record baseline scores for a fixed set of test conversations (10 to 30 is enough to begin)
Block merges when a metric drops below threshold or when the average drops by a set amount compared to baseline
Review failures by category so you can fix the right layer: prompt, retrieval, UI wiring, or model settings
Common mistake: setting one global threshold for everything, then spending hours arguing about a single borderline score. Fix: set different thresholds for different risks, and add a small “allowed drop” window so tiny fluctuations do not fail every run.
If you’re short on time, skip building a large test set and start with your top 10 real user conversations. Run DeepEval on those after any prompt change, and you will still catch most regressions that matter.