Text Evaluation

    Datasets, the Playground, evaluator collections, online evaluation, and batch runs.

    Datasets

    • Upload CSV files, configure input and target columns
    • View and edit rows in-app
    • Or generate a synthetic dataset from a prompt, right from the dashboard

    Playground — Pointwise

    • Single-model generation with full parameter control (temperature, system prompt, max tokens)
    • File attachments — images, video, documents — for multimodal prompts
    • Function tool definitions (JSON schema) and MCP server connections
    • LLM-as-judge scoring per response: relevancy, groundedness, safety, custom metrics

    Playground — Pairwise Comparison

    • Run the same prompt against two models simultaneously
    • LLM-as-judge picks a winner, with reasoning
    • Full comparison history, per project

    Online Evaluation

    Attach an evaluator collection to a tracing project and every production trace is scored automatically as it arrives — no manual step between deploy and quality signal. Configure a polling interval (1–60 minutes) per project; auto-evaluated traces show a green auto badge next to their score.

    Evaluator Collections

    • Pre-built evaluators: relevancy, groundedness, coherence, toxicity, PII detection, custom
    • Three rule types: model-graded, custom prompt, statistical
    • Reusable collections — apply to any project, or set as a project default

    Batch Evaluation

    • Run generation + scoring across an entire dataset in one job
    • Per-row status tracking: pendinggeneratingevaluatingcomplete
    • Regenerate or re-evaluate individual rows without rerunning the whole job
    For programmatic (in-code) scoring with no LLM cost, see Offline Evaluationevalkit.evaluate() runs deterministic checks against tool trajectories and constraints. For synthetic-user testing of a live agent, see Scenario Simulation.

    EvalKit is built by Syntropylabs. Published on PyPI and npm.