Evals, Red Teaming and Test Generation for Agentic Systems

Modular, Lightweight, Dynamic and Async-first

Docs • Website • Community

[!IMPORTANT] Giskard v3 is a fresh rewrite designed for dynamic, multi-turn testing of AI agents. This release drops heavy dependencies for better efficiency while introducing a more powerful AI vulnerability scanner and enhanced RAG evaluation — both now shipping natively in giskard-scan, with no dependency on v2. Only the legacy scan for tabular/ML models remains v2-only. Giskard v2 remains available but is no longer actively maintained. Follow progress → Read the v3 Announcement · Roadmap

Install

pip install giskard           # checks (+ agents, llm, core)
pip install "giskard[scan]"   # + vulnerability / quality scan
pip install "giskard[openai]" # provider SDK for LLM judges / generators

Requires Python 3.12+.

Telemetry: optional aggregated analytics via giskard-core. No prompts or outputs are sent. Opt out with export DO_NOT_TRACK=1 or export GISKARD_TELEMETRY_DISABLED=1 (or the same keys in a .env file in the working directory). Set them before import to skip creating ~/.giskard/id; setting them later still stops further sends. Details: giskard-core README.

Giskard is an open-source Python library for testing and evaluating agentic systems. The v3 architecture is a modular set of focused packages — each carrying only the dependencies it needs — built from scratch to wrap anything: an LLM, a black-box agent, or a multi-step pipeline.

These build on three foundational libraries — giskard-core (shared utilities & telemetry), giskard-llm (provider-agnostic LLM routing), and giskard-agents (agent & workflow orchestration) — which are pulled in automatically and rarely used directly.

Giskard Checks — create and apply evals for testing agents

pip install giskard-checks

Giskard Checks is a lightweight library for creating evaluations (evals) that test LLM-based systems — from simple assertions to LLM-as-judge assessments. Unlike traditional unit tests, evals are designed for non-deterministic outputs where the same input can produce different valid responses.

Use Giskard Checks to:

• Catch regressions — verify your system still behaves correctly after changes
• Validate RAG quality — check if answers are grounded in retrieved context
• Enforce safety rules — ensure outputs conform to your content policies
• Evaluate multi-turn agents — test full conversations, not just single exchanges

Built-in evals include string matching, comparisons, regex, semantic similarity, and LLM-as-judge checks (Groundedness, Conformity, LLMJudge).

Concepts

• Target — your system under test: any sync/async callable (inputs) -> outputs (optionally with trace)
• Scenario — one eval: interactions + checks
• Check — assertion or LLM judge over the trace
• Suite — many scenarios run together

giskard.agents.Generator is an LLM client for workflows/judges — not the same as giskard.checks input generators (LLMGenerator) that synthesize user messages.

Quickstart

import asyncio
from giskard.checks import Scenario, Groundedness


def get_answer(inputs: str) -> str:
    return "Paris"  # replace with your model / agent


async def main() -> None:
    scenario = (
        Scenario("test_france_capital")
        .interact(inputs="What is the capital of France?", outputs=get_answer)
        .check(
            Groundedness(
                name="answer is grounded",
                context="France is in Western Europe. Its capital is Paris.",
            )
        )
    )
    result = await scenario.run()
    result.print_report()


asyncio.run(main())

Groundedness is an LLM judge — install a provider extra (e.g. pip install "giskard[openai]") and set the matching API key. Default model: openai/gpt-4o-mini.

See the full docs for Suites, LLMJudge, multi-turn scenarios, and more.

Giskard Scan — vulnerability scanner for AI agents

pip install "giskard[scan]"   # or: pip install giskard-scan

Giskard Scan is the red-teaming and vulnerability scanning layer for agentic systems. It generates adversarial test suites automatically from a plain-language description of your agent, covering prompt injection, harmful content, stereotypes, misinformation, and more.

Use Giskard Scan to:

• Red-team your agent — automatically generate adversarial inputs across OWASP LLM Top-10 threat categories
• Run prompt-injection probes — built-in dataset of injection payloads ready to use
• Extend with custom generators — pass your own ScenarioGenerator instances to generate_suite, or register them on vulnerability_suite_generator_registry

Quickstart

import asyncio
from giskard.scan import vulnerability_scan


async def my_agent(inputs: str) -> str:
    # Replace with your agent / model call
    return f"Echo: {inputs}"


async def main() -> None:
    await vulnerability_scan(
        target=my_agent,
        description="A customer support chatbot for an e-commerce platform.",
        languages=["en"],
    )


asyncio.run(main())

Scan generators also need an LLM provider extra and API key (same as Checks judges above).

Looking for Giskard v2?

Giskard v2 included Scan (automatic vulnerability detection) and RAGET (RAG evaluation test set generation).

For LLM agents, both are superseded in v3 by giskard-scan: use vulnerability_scan in place of the v2 LLM scan, and quality_scan (with KnowledgeBase) in place of RAGET.

v3 works with ML models too — wrap one as a target and evaluate it with giskard-checks or giskard-scan. What the examples below cover is the v2-only automatic tabular scan — the detector suite that introspects a giskard.Model + giskard.Dataset to auto-detect performance, bias, and robustness issues — along with the giskard.testing ML test suite and the Giskard Hub. These are not planned for v3.

pip install "giskard[llm]>2,<3"

Scan — auto