Giskard v3 is coming!

[mattbit]

We are rewriting the Giskard open-source stack from the ground up 🚀

Why

Giskard v2 tried to cover everything: tabular models, scikit-learn pipelines, pytorch, LLMs APIs… In practice this created a dependency hell, interfaces bent out of shape to accommodate every model type, and a codebase that was hard to maintain, hard to extend, and frankly hard to use.

Giskard v3 takes a step back: the core abstraction is an AI system with one clean API that can wrap anything – an LLM, a classifier, a black-box agent, a multi-step pipeline. Built from scratch on a modular architecture
designed to be extended.

What happens to v2?

Giskard v2 will continue to be available on PyPI (pip install "giskard~=2.0") but will not be maintained. Our focus shifts to v3.

What's in v3

Giskard v3 is structured as a set of focused, complementary packages:

✅ Checks (giskard-checks)
A composable library for testing and evaluating AI agents. Define and run checks against any AI system, from simple assertion to dynamic multi-turn scenarios. For developers testing and shipping advanced agents.

🔍 Agent vulnerability scanner (giskard-scan)
A more powerful, extensible successor of v2 scanner. Focused on red teaming AI agents to detect prompt injections, data leakage, harmful behavior, hallucinations, and more. For developers and security teams who want to red-team their systems.

🗂️ RAG toolkit
An update of our RAG Evaluation Toolkit with new synthetic data evaluation capabilities, including information retrieval benchmarking. For developers evaluating and shipping modern RAG pipelines.

🔗 Hub SDK
Integration with the enterprise Giskard Hub for team collaboration. We are also planning a new SaaS offering for small teams and solo AI developers – more news on this soon!

Each package carries only the dependencies it needs. No more torch or scikit-learn pulled in by default: you only install what you use, and expect it to work across environments and Python versions without surprises.

Under the hood

• giskard-agents: prompt management, agentic workflows, LLM-as-a-judge.
  The same framework powering our enterprise products.

• giskard-core: shared utilities and data structures.

• relais: async streaming pipeline library. Handles the concurrency to make async tasks run fast.

What doesn't change

• Open source, Apache 2.0 licensed.
• The mission: help teams ship AI they can trust.
• The team: same people, new energy!

Get involved!

We're still early. Your feedback can shape what we build.

• Comment and discuss — what's missing, what should work differently, what you'd build with this.
• Follow the progress in the v3 Roadmap
• feature/giskard-v3 — the code is already there. PRs and issues welcome (but reach out to discuss first!)