Response quality analysis
This menu automatically evaluates the quality and safety of the responses generated by an LLM, and monitors risks such as hallucination, toxicity, PII leak, prompt injection, factuality, and answer relevance by category. Use it to diagnose whether "the response content itself is trustworthy and safe" — something that operational metrics such as call volume, cost, and latency alone do not reveal.
Evaluation is performed in a hybrid manner that combines several evaluators such as regular expressions, NLI, and LLM Judge. To control cost, only a sample of all calls is evaluated. Each evaluation item produces a score, and violations are aggregated based on the violation criteria (threshold) of each widget.
From top to bottom, the screen is organized as Summary → Violation trend and Hitmap → Score distribution and model comparison → Time-of-day Hitmap → Violation cases, so you can follow it from grasping the overall status to checking individual violation cases. Set the time range and filters in the top option bar, and then query the widget data.
Summary cards
Cards that show the key quality metrics of the selected time range at a glance.
Evaluation rate
Shows the ratio of calls for which quality evaluation was performed among all LLM calls. The bottom of the card also shows the number of evaluation LLM calls and the total number of calls. (For example, Evaluation LLM calls 170 / 134.8K)
- Evaluation incurs cost, so it runs within the sampling ratio. If the evaluation rate is low, the sample is small, so interpret the ratios in the other widgets with care.
Response quality
Shows the overall quality score of the evaluated responses.
- This is a representative metric that aggregates several evaluation items. Read a downward trend as a signal of quality degradation.
Suspected hallucination rate
Shows the ratio of calls whose hallucination score is at or above the criteria. The bottom of the card also shows the decision criteria and the corresponding number of calls. (For example, score ≥ 0.5 11.4K)
- This indicates how often content that differs from fact is generated in RAG and search-based responses. If the ratio rises, check the context composition or the model.
Factuality
Shows the average level of how well the responses conform to fact (factuality).
- The higher the value, the more faithful the responses are to fact. Viewing it together with the suspected hallucination rate lets you grasp the overall trend of factual distortion.
Toxicity
Shows the ratio of responses evaluated as toxic.
- This monitors how often offensive or harmful expressions that must not be exposed to users occur. If the value spikes, check the actual responses in Violation cases.
Violation trend by category
| Property | Value |
|---|---|
| Chart Type | Stacked area chart |
| Aggregation Unit | Evaluation category |
| Action | Select violation criteria (threshold), toggle legend items |
Shows the trend of violations for each risk category over time as stacked areas. The legend colors correspond to each category (hallucination, toxicity, PII leak, prompt injection, factuality, answer relevance).
-
Use the Violation criteria dropdown at the top right to adjust the score threshold that counts as a violation.
-
If a specific category spikes at a specific point in time, trace the cause by correlating it with deployments or traffic changes.
Operation × category violation rate Hitmap
| Property | Value |
|---|---|
| Chart Type | Hitmap table (rows: operation, columns: category) |
| Aggregation Unit | Operation × category |
| Action | Select violation criteria (threshold) |
Rows are operations and columns are risk categories, and each cell shows the violation rate (%) of that combination. The higher the violation rate, the redder the cell; the lower it is, the greener it is.
-
This widget lets you pinpoint at a glance "which operation is vulnerable to which risk."
-
For example, if the answer relevance column of a specific operation appears red, it means that prompt frequently generates responses that miss the intent of the question.
-
When you adjust the threshold with the Violation criteria dropdown at the top right, the cell colors and values are updated together.
Hallucination score distribution
| Property | Value |
|---|---|
| Chart Type | Histogram (x-axis: score range, y-axis: evaluated calls) |
| Aggregation Unit | Score range (0–10% … 100%) |
| Action | Show values on mouseover |
Shows the distribution of hallucination evaluation scores as the number of calls per score range.
-
If the distribution is concentrated in the low score range (0–10%), it means hallucination is not suspected in most responses.
-
If a thick tail appears in the high score range (especially 100%), a certain volume of responses is suspected of factual distortion, so check the actual cases in Violation cases.
Quality comparison by model
| Property | Value |
|---|---|
| Chart Type | Bar chart (x-axis: model) |
| Aggregation Unit | Model |
| Action | Show values on mouseover |
Compares the response quality score of each model as bars.
-
A model with a low quality score may cause more violations under the same workload.
-
Review it together with the cost and latency metrics to decide whether to switch to a model with better efficiency relative to quality.
Factuality score distribution
| Property | Value |
|---|---|
| Chart Type | Histogram (x-axis: score range, y-axis: evaluated calls) |
| Aggregation Unit | Score range (0–10% … 100%) |
| Action | Show values on mouseover |
Shows the distribution of factuality evaluation scores as the number of calls per score range.
-
If the distribution is concentrated in the high score range (especially 100%), it means most responses are faithful to fact.
-
If the proportion of the low score range (0–10%) is thick, a certain volume of responses differs from fact, so interpret it together with the hallucination score distribution.
Violation Hitmap by time of day
| Property | Value |
|---|---|
| Chart Type | Hitmap (day of week × time of day) |
| Aggregation Unit | Day of week (Sun–Sat) × hour (0–23) |
| Action | Select violation criteria (threshold) |
Shows the density of violations on a grid of days of the week (vertical) and hours (horizontal). The darker the cell, the more violations are concentrated in that time slot.
-
If violations cluster in specific days or hours (for example, weekday afternoons), check the correlation with traffic load or specific user patterns.
-
Use the Violation criteria dropdown at the top right to adjust the threshold.
Violation cases
| Property | Value |
|---|---|
| Chart Type | Table |
| Action | Select violation criteria (threshold), expand rows |
Lists the individual calls determined to be violations by the evaluation, one row per call. This widget is for checking actual cases directly rather than statistics.
| Column | Description |
|---|---|
Time | Time the call occurred |
Operation | Operation in which the violation occurred |
Prompt | Summary of the input and the output |
Model | Name of the LLM model used |
Provider | LLM provider |
Token | Total number of tokens |
Latency | Time from request start to response completion |
Cost | Cost of the call |
-
Expanding a row shows the original input and response along with the score of each evaluation item in the
Evalarea. (Hallucination, Factuality, Relevance, Toxicity, Prompt Injection, PII Leak, URL Scan, Judge) -
Use the Violation criteria dropdown at the top right to adjust the threshold for marking violations.
Reference
Evaluation categories
The category names on the screen correspond to the evaluation items in the Eval area as follows.
| Category | Eval Item | Description |
|---|---|---|
| Hallucination | Hallucination | Degree to which content not in the context was generated |
| Factuality | Factuality | Degree to which the response conforms to fact |
| Answer relevance | Relevance | Degree to which the response matches the intent of the question |
| Toxicity | Toxicity | Degree to which offensive or harmful expressions are included |
| Prompt injection | Prompt Injection | Attempts to hijack or bypass the system prompt |
| PII leak | PII Leak | Degree of personally identifiable information exposure |
| - | URL Scan | Whether URLs included in the response are risky |
| - | Judge | Overall decision score of the LLM Judge |
Violation criteria (threshold)
-
For the violation trend by category, violation rate Hitmap, violation Hitmap by time of day, and violation cases widgets, you can select the Violation criteria (score threshold) at the top right.
-
When you adjust the threshold, the criteria by which that widget counts violations changes, and the data is updated together.
-
Evaluation is performed on a sample of all calls, so the violation rate is calculated based on the evaluated calls.