giskard-llm

Overview

Distinct from functional eval frameworks (Promptfoo, DeepEval, Ragas), Giskard's value (per gk-gh) is adversarial test generation - auto-generates inputs designed to break LLMs along documented vulnerability dimensions, then reports findings in a triageable HTML report.

Important version note (2026-05-06): per gk-gh, "Giskard v2 is no longer actively maintained. The current v3 focus is on giskard-checks for evaluations, while vulnerability scanning and RAG evaluation still rely on Giskard v2." This skill targets v2 LLM scanning; pin >2,<3 per the install command.

When to use

• The team needs a red-team / adversarial pass before shipping an LLM feature.
• Functional evals (Promptfoo / DeepEval / Ragas) pass but the team wants to surface vulnerabilities beyond the test corpus.
• Compliance-driven assurance needs reportable evidence of hallucination / harmful-content / prompt-injection coverage.
• Product owners need an HTML report (not raw test output) to triage and sign off.

Step 1 - Install

Per gk-gh (v2-pinned):

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

The [llm] extra pulls in the LLM scan dependencies.

Step 2 - Wrap your model

Per gk-gh (verbatim quickstart):

import giskard
import pandas as pd

def model_predict(df: pd.DataFrame):
    """The function takes a DataFrame and must return a list of outputs (one per row)."""
    return [my_llm_chain.run({"query": question}) for question in df["question"]]

giskard_model = giskard.Model(
    model=model_predict,
    model_type="text_generation",
    name="My LLM Application",
    description="A question answering assistant",
    feature_names=["question"],
)

The description field steers Giskard's adversarial generator - make it specific (e.g., "A question answering assistant for medical guidance" vs the generic "A QA assistant"). Better description ⇒ better-targeted adversarial inputs.

Step 3 - Run the scan

scan_results = giskard.scan(giskard_model)
display(scan_results)

(Per gk-gh.)

In a Jupyter notebook this renders the report inline. To export:

scan_results.to_html("giskard-report.html")

Step 4 - Vulnerability detector catalog

Per gk-gh the v2 LLM scan covers these detector categories:

Detector selection (subset run):

scan_results = giskard.scan(
    giskard_model,
    only=["hallucination", "prompt_injection"],
)

(Per the scan() API; consult docs.giskard.ai for the exact parameter list per Giskard release.)

Step 5 - Convert findings to test suites

After a scan surfaces issues, Giskard can synthesize tests for regression coverage:

test_suite = scan_results.generate_test_suite("My LLM Test Suite")
test_suite.run()

This produces deterministic regression tests from the failing adversarial prompts found by the scan - re-run on every PR to prevent regression on previously surfaced vulnerabilities.

Step 6 - CI integration

Giskard does not ship a first-party CI action; pattern:

python -m my_giskard_scan_script  # produces giskard-report.html

Then upload as a CI artifact:

- uses: actions/upload-artifact@v4
  with: { name: giskard-report, path: giskard-report.html }

For PR-blocking gating, parse scan_results for severity and fail CI if any critical findings appear:

critical = [issue for issue in scan_results.issues if issue.level == "major"]
if critical:
    sys.exit(1)

Anti-patterns

Limitations

• v2 LLM scanning is in maintenance-only mode (per gk-gh); v3 is forming around giskard-checks. Track upstream before greenlighting new investment.
• Adversarial generation is non-deterministic - use random seeds when available + pin Giskard version.
• LLM-as-judge cost: scans invoke a judge model many times; budget for cost spikes when scanning new models.
• Limited to text_generation and text_classification model types in v2; multi-modal scanning lives elsewhere in the Giskard ecosystem.

References

• gk-gh - repository, install, quickstart, detector list
• gk-docs - full documentation
• promptfoo-evaluation, deepeval-evaluation, ragas-evaluation - functional eval sister tools (use Giskard for adversarial coverage on top)