LLM-as-a-judge

Automatically evaluate production LLM traces for quality and safety — no code changes required

LLM-as-a-judge evaluates your LLM responses automatically as they appear in OpenLIT traces. You configure it once in the UI and it continuously scores every sampled response for hallucination, bias, toxicity, and other quality signals — without any changes to your application code.

LLM-as-a-judge requires that your application already sends traces to OpenLIT. If you haven’t set up tracing yet, see the tracing guide first.

How it works

When a trace arrives in OpenLIT, the evaluation engine extracts the LLM prompt and response, then sends them to a judge LLM with a structured scoring prompt. The judge returns a JSON object containing a numeric score (0–1), a classification, an explanation, and a pass/fail verdict. OpenLIT stores this result alongside the original trace so you can view it in the trace detail panel and query it in dashboards.

The full evaluation loop is:

1. Your application sends a trace to OpenLIT (via the SDK or OpenTelemetry)
2. OpenLIT samples the trace according to your configured frequency
3. The judge LLM scores the response across the enabled evaluation criteria
4. Results are stored and surface in the trace detail view

Supported evaluation criteria

LLM-as-a-judge can evaluate responses across three categories. Each category is broken into specific classification types that the judge LLM uses when reporting results.

Hallucination

Detects when an LLM response contradicts or diverges from the provided context.

Classification	Description
factual_inaccuracy	The response states facts that contradict the source context
nonsensical_response	The response is irrelevant to the prompt or context
gibberish	The response contains incoherent or meaningless text
contradiction	The response conflicts with itself or with the context

Bias

Detects language that reflects unfair assumptions or stereotypes about people.

Classification	Description
sexual_orientation	Assumptions about an individual’s sexual preferences
age	Stereotypes based on a person’s age
disability	Assumptions about competence due to disability
physical_appearance	Judgments based on physical characteristics
religion	Prejudices connected to religious beliefs
pregnancy_status	Bias toward pregnant individuals or parents
marital_status	Assumptions based on relationship status
nationality / location	Stereotypes based on country or place of origin
gender	Assumptions about capabilities based on gender
ethnicity	Generalizations based on racial or ethnic background
socioeconomic_status	Bias based on economic or social position

Toxicity

Detects harmful or aggressive language in LLM responses.

Classification	Description
threat	Language that threatens harm or negative consequences
dismissive	Language that belittles or disregards the user
hate	Hateful or derogatory language
mockery	Sarcastic or ridiculing language
personal_attack	Direct attacks on a person’s character or intelligence

Configure LLM-as-a-judge

1. Open the Evaluations page

   In the OpenLIT UI, navigate to Evaluations in the left sidebar. This page lists all configured evaluation rules and shows recent evaluation results.

2. Create a new evaluation rule

   Click Create rule. Configure the following settings:

   • Name — a descriptive label for this evaluation rule
   • Judge model — the LLM provider and model to use as the judge (e.g., openai / gpt-4o-mini)
   • API key — the API key for the judge LLM provider
   • Evaluation criteria — select one or more of: Hallucination, Bias, Toxicity, or All
   • Threshold score — the cutoff score (0–1) above which a response is marked as failing. Defaults to 0.5
   • Sampling rate — the percentage of traces to evaluate. Lower rates reduce costs on high-volume applications
3. Enable the rule

   Toggle the rule to Active. OpenLIT will begin evaluating incoming traces according to the configured sampling rate.

View evaluation results

Evaluation results appear in two places in the OpenLIT UI:

Trace detail panel — open any trace and scroll to the Evaluations section to see the score, verdict, classification, and explanation for each evaluation that ran against that trace.

Evaluations dashboard — the main Evaluations page shows aggregate evaluation results across all traces, including score distributions, verdict breakdowns, and trends over time. You can filter by evaluation type, time range, and model.

If you see a high proportion of hallucination verdicts, cross-reference those traces with the prompts and context documents your application provides. A common cause is missing or stale context in RAG pipelines.

Evaluation result fields

Every evaluation result contains the following fields:

Field	Type	Description
score	float (0–1)	Numeric confidence that the issue is present. Closer to 1 means higher likelihood
evaluation	string	The evaluation type: hallucination, bias_detection, or toxicity_detection
classification	string	The specific sub-category detected (e.g., factual_inaccuracy, age)
explanation	string	A short sentence explaining why the issue was detected
verdict	string	yes if the score exceeds the threshold, no otherwise

---

Programmatic evaluations

Run evaluators directly in your Python code for CI/CD quality gates

Tracing

Set up automatic LLM tracing as a prerequisite for LLM-as-a-judge