Skip to main content
You are reading the docs of Galtea 5.15. Read this page in the latest version.

What are Metrics?

Metrics define how Galtea scores your product’s outputs during evaluations. Each metric captures one specific quality dimension — factual accuracy, tone, security resilience, or a custom criterion you define.
Metrics are organization-wide and can be reused across multiple products. You can link metrics to Specifications for structured evaluation workflows.
You can create, view and manage your metrics on the Galtea dashboard or programmatically using the Galtea SDK.

Two Families of Metrics

  • Deterministic metrics apply rule-based logic — string matching, numerical checks, bounding box overlap. Results are consistent and reproducible. Built-in examples: BLEU, ROUGE, Text Similarity, Tool Correctness. You can also define your own using the SDK’s CustomScoreEvaluationMetric class — see our tutorial.
  • Non-deterministic metrics use an LLM-as-a-judge to assess open-ended qualities like factual accuracy, misuse resilience, and task completion. Galtea’s judges are human-aligned and optimized for each evaluation type. You can also create your own AI Evaluation metrics with a custom scoring rubric.

How Metrics Are Evaluated

Every metric has an Evaluation Type that determines how scoring happens:

Self-Hosted

Your own logic runs locally; only the score is uploaded.

AI Evaluation

An LLM scores responses using a rubric you define.

Human Evaluation

Human annotators score responses using defined criteria.

Revisions and Analytics

Creating a revision marks the previous metric revision as Legacy. Legacy and analytics inclusion are separate states. A legacy revision keeps contributing to scores and coverage unless you exclude it. After you save a revision, the dashboard shows how many results and products the earlier revision affects. A user limited to the products their user groups grant sees counts that cover only those products, but the exclusion applies to every product. The default keeps those results in analytics. Select Exclude from analytics only when you want both the scores and coverage to stop counting them. You can reverse this decision from the metric’s actions menu. Exclusion does not delete or change any evaluation. Each evaluation keeps its score, reason, and metric revision. The metric list hides excluded revisions by default. To reveal them, open Filters and turn on Show excluded from analytics. To inspect earlier revisions, open Filters and turn on Include legacy. Add the Analytics column from the Columns menu to see whether each revision is Included or Excluded. When a product has excluded results, the Details view of its Analytics page states how many results are outside the reported numbers. The link opens the evaluation list with Show excluded from analytics active. Excluded revisions do not appear in the Analytics metric filter because they contribute no scores to that page. The page’s Last Successful date still counts their evaluations, because it reports when an evaluation last ran.

Pass threshold

Every metric has a pass threshold: the score, from 0 to 1, at or above which an evaluation passes. When you do not set one, it is 0.5. Galtea’s own metrics use 0.5. An evaluation with no reported score has no result, so it neither passes nor fails. You set the threshold when you create the metric. It cannot change afterwards, so to use another one, create a revision of the metric. In the dashboard, the threshold is a percent in Advanced Options of the create form, prefilled with 50, and it accepts decimals. A revision starts with the threshold of the revision it comes from. Score pills are green at or above the threshold and red below it.

Applicability check

Some metrics apply to few sessions. A jailbreak metric, for example, applies only to the sessions where a user tries a jailbreak. If the judge scores every session, it scores many sessions the metric was not made for. Those sessions lower the pass rate or pass for free, and each judge call costs credits. The applicability check stops this. Before the judge runs, a decision model answers one Yes/No question: does the conversation contain what this metric evaluates? When the probability that it does is below the metric’s applicability threshold, the evaluation is skipped. The check runs before a JSON Field Match metric reads a test case’s field rules, so a session it skips never reports an error in them. The decision model sees each rule as raw JSON. A monitor runs the check for each production session, after the data check and before the judge. See How a session is evaluated for the phases and the skip rules.

Settings

Two settings control the check. Both are fixed when you create the metric. To change one, create a revision of the metric. A revision starts with the values of the revision it comes from, when the check runs for both revisions.
  • Run the check on (applicability_scope) sets which sessions get the check. The options are PRODUCTION (production sessions), DEVELOPMENT (development sessions), ALL and NONE (no check). The default is PRODUCTION. Development is off by default, because a skip inside a regression suite can hide a test that should run.
  • Minimum certainty (applicability_threshold) is the lowest probability, from 0 to 1, that the metric applies to a session at which the metric is judged. A higher value skips more sessions. The default is 0.5. A session at the threshold is judged. A session below it is skipped.
In the dashboard, both settings are in Advanced Options of the create form. Setting Run the check on to Off hides Minimum certainty. The form hides both settings for a metric that has no check. The dashboard shows Minimum certainty as a percent from 0 to 100, prefilled with 50. The SDK and the API take a probability from 0 to 1. The check covers four kinds of metric: AI Evaluation, Classifier, deterministic, and the DeepEval-based metrics of Galtea. A Self-Hosted, Human Evaluation or G-Eval (deprecated) metric has no check: the scope is NONE, and the API refuses any other value. A metric that scores traces has no check either, because the decision model never reads traces. Its scope is also NONE, and the API refuses any other scope and any threshold. For a classifier metric, the fields it scores are the placeholders of its question. Metrics that existed before the check got PRODUCTION when the check covers their kind and they do not score traces. The others kept NONE.
The check runs only where your deployment enables it. Where it is off, every session is judged and the two settings have no effect. On an on-premise deployment, set the environment variable METRIC_APPLICABILITY_CHECK of the API to off (the default), shadow or enforce, and set VITE_METRIC_APPLICABILITY_CHECK_ENABLED of the dashboard to true to show the two settings. The check needs the LLM gateway (LiteLLM) to route to the decision model. Without that route, every session is judged as if there were no check. In shadow mode, the check runs and writes what it would skip to the evaluator logs, but every session is still judged. Use it to measure the skip rate of a metric before you enforce.

What the check reads

The decision model reads the input and the actual_output of each turn, the other fields the metric scores (for example retrieval_context, tools_used or expected_output), and the name and description of the metric. It also reads the metric’s criteria: the judge prompt, or the question of a classifier metric. For an AI Evaluation metric whose judge prompt names a field, such as tool calls, that field counts as scored. It never reads traces: one turn’s traces can be longer than everything the decision model reads at once. It has no size limit of its own. When a session is too long for the decision model, the metric is judged. Galtea sends this data to the provider of the deployment’s decision model. See Data processing. The check never blocks an evaluation by failing. The metric is judged as if there were no check when:
  • the decision model is not available, or it does not answer in time,
  • the conversation is too long for the decision model,
  • the probability is below the threshold, but the conversation is longer than about 2,000 tokens. The answer is less reliable on a long conversation, so Galtea does not skip on it.
The check costs no credits.

A skipped evaluation

A skipped evaluation has no score and no judge call, and it costs no credits. Its error message gives the reason, for example: “The metric does not apply to this session: the probability that it applies is 12%, below its applicability threshold of 50%.” A retry would ask the same question about the same session, so canRetry is false. When the skip is a session’s latest evaluation in its metric family, the session stays out of that family’s count in the monitor’s alert rules, unless a reviewer scored the evaluation. See Alert rules. Analytics still count the skip in its skipped total. For a metric that applies to few sessions, raise the monitor’s sampling percentage, so that enough sessions reach the check. To correct a skip, use one of these:
  • A reviewer scores the evaluation. The human score turns the skipped row into a scored one.
  • Create a revision of the metric with Off as the scope, and evaluate the session again with it.

Optimization review

Optimize writes an improved judge prompt into a new revision named <name> (Optimized). While it runs, that revision’s page shows its progress, and you can leave the page. Unlike other revisions, it does not mark the original Legacy. When the optimizer returns a different prompt, the revision shows Ready for review, and its page shows the changes and both prompts.
  • Apply optimization marks the original revision Legacy. The optimized revision takes the original’s name, description, and tags, and joins its specifications and user groups. Monitors use it from their next run.
  • Keep original leaves the original unchanged and marks the optimized revision Not applied. You can then optimize again.
The optimizer judges each prompt at the metric’s pass threshold, and the optimized revision keeps that threshold. From 30 annotated evaluations, a run keeps some of them as a validation set. The optimizer does not train on the validation set. An evaluation kept for validation stays out of training in later attempts of the same metric. The Optimize confirmation shows how many evaluations go to the training set and how many to the validation set. A user limited to the products their user groups grant sees counts that cover only those products. With fewer than 30 annotated evaluations, there is no validation set unless an earlier attempt kept one. Optimize is disabled when the evaluations kept for validation leave too few to train on. Expand Advanced Options in the Optimize confirmation to narrow a run. Strategies sets which rewrites the optimizer tries: Fix false passes, Fix false failures, Add a scoring rubric, Simplify the prompt and Broaden the wording. Each step asks for one rewrite per selected strategy and keeps the best one. Maximum steps sets how many steps the run may take, from 1 to 10, and defaults to 10. A run stops earlier when the prompt stops improving. With every strategy selected and 10 steps, the run is the same as one without options. The review card, the No improvement notice and the page of a declined or applied revision show the steps the run used out of its maximum and the strategies it ran. An attempt that did not report them shows Not reported. The review compares the original and the optimized prompt on both sets. For each prompt, it shows how many of the evaluations that both prompts scored agree with your annotation, and the alignment when it is defined. A declined or applied revision keeps these results on its page. The validation result is evidence, not a guarantee: a small validation set gives a rough estimate. Only the active revision of a metric can be optimized. Through the API, call POST /metrics/{id}/activate-optimization or POST /metrics/{id}/decline-optimization with the optimized revision’s id.

Evaluator model retirement

The evaluator model of a metric is its LLM judge, or the classifier model of a Classifier metric. When an evaluator model is retiring, you get an email up to 30 days before the retirement date. It names the affected metrics and the evaluator models you can use instead. It offers only models of the same kind as the retiring one: judge models for a judge model, classifier models for a classifier model. The email goes to the organization owners and admins, and to the creator of each affected metric. A metric’s evaluator model cannot change after you create it. For your own metrics, create a new revision with another evaluator model. Galtea maintains its own metrics, so you cannot revise them. If your organization used one in the last 90 days, its owners and admins get the email too. Galtea may replace one of its metrics with a new revision on another evaluator model. When it does, the email names the new revision and links to it. Use that revision to keep evaluating. When there is no new revision, create your own metric with one of the listed evaluator models to keep evaluating after the retirement date. If the evaluator model becomes unavailable to Galtea on or after the retirement date, and stays unavailable for six hours, the same recipients get one email naming the metrics that stopped scoring.

Available Metrics

The following table lists the default metrics available in the Galtea platform. You can also create custom metrics or generate metrics from specifications.

SDK Integration

Metrics Service SDK

Manage metrics using the Python SDK

Metric Properties

Text
required
The name of the metric. Example: “Factual Accuracy”
Text
A brief description of what the metric evaluates.
Text
The model used to score the metric. Does not apply to deterministic metrics. For a Classifier metric, it is a classifier model, picked in the form or sent as evaluator_model_name. Example: “GPT-4.1”
Number
The score, from 0 to 1, at or above which an evaluation of this metric passes. Defaults to 0.5. It is fixed when you create the metric: create a revision to change it. In the SDK it is metric.pass_threshold. Example: 0.7
Enum
Which sessions get the applicability check: PRODUCTION, DEVELOPMENT, ALL or NONE. Defaults to PRODUCTION for a metric the check covers, and to NONE for a Self-Hosted, Human Evaluation or G-Eval (deprecated) metric. A metric that scores traces never gets the check: it defaults to NONE and refuses any other value. It is fixed when you create the metric: create a revision to change it. In the SDK it is metric.applicability_scope. Example: “PRODUCTION”
Number
The lowest probability, from 0 to 1, that the metric applies to a session at which the metric is judged. Defaults to 0.5, and reads None in the SDK until you set it. A metric with no check refuses it: a Self-Hosted, Human Evaluation or G-Eval (deprecated) metric, or one that scores traces. It is fixed when you create the metric: create a revision to change it. In the SDK it is metric.applicability_threshold. Example: 0.3
Text List
Tags for categorization. Example: [“RAG”, “Conversational”]
Enum
required
How outputs are scored: AI Evaluation, Human Evaluation, or Self-Hosted. See Evaluation Types for full details.
List[String]
User groups for Human Evaluation metrics. Controls which annotators can score evaluations for this metric.
List[Enum]
required
The data fields available to the evaluator. See Evaluation Parameters for the full reference. Not applicable for Self-Hosted metrics. For a Classifier metric, the platform derives them from the placeholders in the question.
Object
The one Yes/No or Scale question a classifier metric asks. Only for the classifier source. See Classifier.

Concepts overview

How Galtea’s concepts connect — diagram + per-entity quick reference.

Evaluation

The assessment of an evaluation using a specific metric’s criteria

Create Custom Metrics

Write your own judge prompt and scoring rubric.