Giskard v3 is here!

[kevinmessiaen]

🎉 Giskard v3 is now available. 🎉

This release rebuilds Giskard around a simpler goal: help developers test, evaluate, and red-team AI agents with focused Python packages and a lightweight, async-first architecture.

Build repeatable agent evaluations

Use Giskard Checks to describe single-turn or multi-turn scenarios, run them against synchronous or asynchronous targets, and evaluate the resulting trace with deterministic checks or LLM judges.

Checks can cover exact and semantic matching, groundedness, conformity, JSON validity, readability, custom assertions, and more. Scenarios can be grouped into suites, run in parallel, and exported as JUnit reports for CI/CD.

Generate security and quality tests

Use Giskard Scan and vulnerability_scan to generate adversarial scenarios for risks such as prompt injection, jailbreaks, harmful content, stereotypes, and misinformation.

For RAG systems, quality_scan with a knowledge base generates and evaluates knowledge-base quality scenarios. Generated suites can also be replayed in CI.

Try it out

The base package installs Giskard Checks:

pip install giskard

Add Scan for vulnerability and quality scanning, and install the provider extra required by any LLM-backed checks or generators:

pip install "giskard[scan]"
pip install "giskard[openai]"

See the Checks installation guide or Scan installation guide for provider setup.

⚠️ Giskard v3 requires Python 3.12 or newer.

A new architecture

v3 is organized as five focused libraries:

• giskard-checks for agent testing and evaluation
• giskard-scan for vulnerability and RAG-quality scanning
• giskard-agents for agent and workflow orchestration
• giskard-llm for provider-agnostic LLM routing
• giskard-core for shared utilities

Each package carries only the dependencies it needs, so projects can install the capabilities they use without bringing in the full stack.

Migrating from v2

v3 is a full rewrite rather than an in-place API upgrade, check out the [v2-to-v3 migration] guide(https://docs.giskard.ai/oss/migrate-from-v2).

The automatic tabular/ML scan and the legacy @giskard.testing` suite remain v2-only. If you rely on those workflows, keep that part of your project on v2:

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

Giskard v2 remains available but is no longer actively maintained.

Start testing

• Browse all OSS documentation
• View the v3.0.0 release
• Browse the source or open an issue
• Join the community on Discord