Traigent eval choose metric

Traigent Choose Metric

When to Use

Use this skill before configuring @traigent.optimize() when the user has not yet pinned what "good" means.

• Prefer plain objectives=["accuracy", "cost"] lists unless the project already needs weighted objective schemas.
• Use built-in metric names when they match the job: accuracy, success_rate, error_rate, avg_output_length, cost, latency.
• Use custom metric_functions when the task has checkable domain logic that a built-in metric cannot express.
• If the key property is a must-not-violate safety condition, treat it as a gate or constraint, not as an ordinary objective.

The metric interview

Ask these six questions and write down the answers before coding:

1. What does "good" mean in product terms: correct answer, better rank, valid JSON, safe refusal, lower cost, lower latency, or higher reliability?
2. Is there checkable ground truth, or does the team need a rubric/judge?
3. Is the unit one input-output example, a multi-turn conversation, or an agent trace with tools?
4. Which failure is unacceptable even if the average score improves?
5. What budget matters: dollars, latency, tokens, calls, tool invocations, or review time?
6. Who consumes the score: optimizer, CI gate, release owner, customer report, or debugging workflow?

Convert the answers into one primary metric, optional secondary objectives, and any hard gates.

The no-gold track (Q2 = neither)

If Q2 comes back with no checkable ground truth at all — subjective or generative quality (summaries, chat, writing, open-ended agents) with no labels to compare against — do not stall and do not force an exact-match metric. Route onto the judge track:

1. Write a short rubric for what "good" means in product terms (from Q1), with 2-3 anchored score levels.
2. Build the metric from traigent-eval-build's "LLM judge with rubric, strict parse, and cost guardrails" template — the judge score is the primary quality KPI (label it for what it is, e.g. judge_quality; the accuracy-labeled default does not apply here — say so).
3. traigent-eval-audit is mandatory on this track, not optional: the judge now owns the entire quality signal, so its agreement/stability audit is the only thing standing between you and optimizing noise.
4. Budget the judge itself: every trial example costs a judge call on top of the agent call. Flag judge cost as its own line in the run budget (and consider cost as a secondary objective so expensive judge-pleasing configs don't win by default).

Measure-type grounding

Use this vocabulary in prose and tables. Do not import these names.

Measure type	What it asks	Common evaluation method
sanity_check	Does the wiring work and return parseable outputs?	deterministic
accuracy	Does the output match a known label, answer, or test result?	deterministic or statistical
quality	Is the answer useful, complete, grounded, or well-written?	llm_based or hybrid
latency	How long does the call or workflow take?	deterministic
safety	Did the system avoid a prohibited action or output?	deterministic or hybrid
efficiency	Did the system use fewer tokens, calls, tools, or dollars?	deterministic
reliability	Does the behavior remain stable across repeats or noisy inputs?	statistical or hybrid

The built-in latency objective binds to the bare key latency, reported in milliseconds on SDKs after 0.22.0 (see version-matrix: latency-unit).

Evaluation methods:

Method	Use when
deterministic	Exact match, normalized match, schema validation, tests, cost, latency, or tool-call rules can be checked directly.
llm_based	A rubric is required and deterministic labels are insufficient. Label the result as a judge score. LLM-based evaluators are subject to service-side evaluator-quality auditing (ACET) after runs complete — the optimizer tensor is used to retroactively assess the judge, no new gold collection required. Favor a verifiable anchor (execution-match, unit-test pass, MCQ exact match) where the task allows, as it enables honest confidence from the service audit.
statistical	Repeated runs, variance, agreement, or confidence intervals matter.
hybrid	A deterministic gate should run before a judge or repeated-sample score.

From answers to objectives

Start with built-ins when possible:

import litellm
import traigent

def prompt_model(prompt: str, *, model: str, temperature: float = 0.0) -> str:
    response = litellm.completion(
        model=model,
        temperature=temperature,
        messages=[{"role": "user", "content": prompt}],
    )
    return response.choices[0].message.content or ""

@traigent.optimize(
    objectives=["accuracy", "cost"],
    configuration_space={
        "model": ["gpt-4o-mini", "gpt-4o"],
        "temperature": [0.0, 0.3],
    },
)
def answer(question: str) -> str:
    cfg = traigent.get_config()
    return prompt_model(question, model=cfg["model"], temperature=cfg["temperature"])

Use custom metric functions when the domain has a checkable rule. Default to an accuracy-labeled primary quality objective when correctness or answer quality applies, either as a built-in objective or as a metric_functions key. The label does not have to be the literal string accuracy: an accuracy-semantic name that carries a qualifier (label_accuracy, exec_accuracy) satisfies the rule and is preferred when it says which accuracy is measured. Name the primary quality metric after the product concept only when accuracy semantically does not fit the problem, such as ranking quality, generation quality, schema validity, or latency-only tuning; note why accuracy was skipped, and include that name in objectives only if the optimizer should trade off against it.

In the schema-only example below, valid_schema is the product-concept KPI because output format validity is the primary target; add an accuracy metric too if answer correctness also matters.

import json

import traigent
from traigent.api.decorators import EvaluationOptions

def valid_schema_metric(output, expected, input_data) -> float:
    try:
        data = json.loads(output)
    except json.JSONDecodeError:
        return 0.0
    required = {"invoice_id", "amount_due", "due_date"}
    return 1.0 if required.issubset(data) else 0.0

@traigent.optimize(
    evaluation=EvaluationOptions(
        eval_dataset="eval/invoices.jsonl",
        metric_functions={"valid_schema": valid_schema_metric},
    ),
    # Schema validity is the primary target here (accuracy skipped: format-only
    # task). Add an accuracy metric if answer correctness also matters.
    objectives=["valid_schema", "cost"],
    configuration_space={"temperature": [0.0, 0.2]},
)
def extract_invoice(text: str) -> str:
    cfg = traigent.get_config()
    return prompt_model(
        f"Extract invoice_id, amount_due, and due_date as JSON from:\n{text}",
        model="gpt-4o-mini",
        temperature=cfg["temperature"],
    )

If you need weighted objective schemas, verify the exact ObjectiveSchema and ObjectiveDefinition import path against the installed SDK first. The public examples should prefer plain objective lists unless weights are essential.

Multi-objective patterns

Use accuracy + cost as the default two-objective pattern for LLM tasks. It keeps the baseline quality target visible while discouraging expensive wins that are not worth the difference.

Pattern	Use
["accuracy"]	Early correctness tuning with a fixed budget outside the objective.
["accuracy", "cost"]	Default for answer quality where spend matters.
["accuracy", "latency"]	User-facing online flows where response time matters (bare latency key, in ms on SDKs after 0.22.0 — see version-matrix: latency-unit).
["valid_schema", "accuracy", "cost"]	Extraction tasks with machine-checkable output format.
["success_rate", "cost"]	Tool or agent workflows where execution success is the main signal.

Treat safety properties as constraints or gates when a violation is unacceptable. Do not let a safety score be traded away for better average accuracy or lower cost. Hand off to the relevant gate/release policy skill when the user needs approval semantics.

Claim scope

• Metric choice determines what the optimizer can see; it is not proof that unmeasured properties improved.
• Built-in metrics are useful only when their definitions match the product claim.
• LLM-based quality scores are model opinions under a rubric. Label them that way.
• Multi-objective results depend on the dataset, objective names, and run budget used for that optimization.

See Also

• traigent-eval-build - next step after choosing the metric
• traigent Step 3.5 - lightweight evaluator sanity gate (known-good/known-bad assertion before first real run)
• traigent-setup-decorator - wiring objectives and evaluation options into the decorator
• traigent-analyze-results - reading trial metrics and reporting outcomes

<!-- Reserved: managed longitudinal-guidance region. Step-level edits must not write here. --> <!-- SLOW_UPDATE --> <!-- /SLOW_UPDATE --> <!-- INTERACTION_POLICY v1 (synced — do not edit inline; edit docs/shared/interaction-policy.v1.md) -->

Traigent Interaction Policy

Track an interaction profile and adapt to it. Persona (stable): control=delegate|guided|inspect, expertise=se|ds|unknown. Mood (this session): pace=execute|balanced|explore. Default when unknown: guided,se,balanced. Infer from explicit user statements first, then recent behavior; an explicit correction wins immediately. Never store or send this profile anywhere by default.

Fetch the live profile (when available)

At session or skill start, if a configured Traigent client is available, seed the profile from the backend with the skill name:

policy = None
try: policy = await client.get_interaction_policy(skill="<this skill>")
except Exception: pass

Treat the returned profile as the STARTING seed: its control/expertise/pace axes plus question_budget, options_max, and jargon_level replace the static defaults below. Explicit user corrections in-conversation ALWAYS override the seed. If the call is unavailable or fallback_policy="static_v1", simply use the static defaults below; the SDK already fails soft.

• Always be concise.
• Match terminology to expertise. For se: plain engineering words; define each Traigent or statistics term once in plain language (no Bayesian / variance-decomposition / Pareto jargon unless asked). For ds: compact optimization and statistical terms are fine.
• Presenting options: show at most 3, mark exactly one Recommended, and give one short persona-appropriate trade-off per option.
• Autonomy. For delegate or execute: pick the recommended reversible action and proceed, asking only at hard gates. For guided: offer options with a recommendation at the key decisions. For inspect or explore: give brief rationale or evidence before asking, and ask before branch choices.
• Hard gates — always confirm regardless of persona: paid or provider model calls, sending data or private content off the machine, destructive edits, decisions the Traigent service is meant to return, and any missing fact the step truly requires.
• Always end by recommending the next Traigent skill or action to take.
• Never weaken Traigent safety: dry-run before any paid run; get explicit approval before real cost or before any data leaves the machine; treat service-returned plans and next steps as authoritative. Never put the persona profile or any private content into telemetry, run metadata, experiment names, logs, or provenance files.

<!-- /INTERACTION_POLICY v1 -->