Klarity is a generative AI toolkit designed for inspecting and debugging AI decision-making processes. It targets AI developers and researchers seeking to understand model behavior, mitigate errors, and improve model reliability through automated explainability, uncertainty analysis, and multi-modal support. The toolkit offers structured insights into model confidence, reasoning patterns, and visual attention, enabling more robust AI systems.

How It Works

Klarity employs a multi-faceted approach to AI explainability. It quantifies model confidence using raw entropy and semantic similarity metrics, analyzes step-by-step reasoning patterns extracted from model outputs, and visualizes visual attention in Vision-Language Models (VLMs). These analyses are synthesized into structured JSON outputs and AI-powered reports, providing actionable insights for debugging and improvement. The core advantage lies in its integrated approach, combining uncertainty, reasoning, and visual attention for a holistic view of model decision-making.

Quick Start & Requirements

• Install via pip: pip install git+https://github.com/klara-research/klarity.git
• Requires Hugging Face Transformers, vLLM, or Together AI API for analysis.
• Specific models require GPU acceleration (e.g., LlavaOnevisionForConditionalGeneration, DeepSeek-R1-Distill-Qwen-7B).
• API keys are needed for certain insight models (e.g., Together AI).
• Official documentation and examples are available on GitHub.

Highlighted Details

• Achieves 80%+ accuracy in hallucination detection using entropy + judge LLM metrics.
• Supports integration for VLM and visual attention monitoring.
• Enables analysis of reasoning model CoTs, entropy, and RL dataset improvement.
• Provides detailed JSON outputs for VLM, Reasoning, and Entropy analyses.

Maintenance & Community

• Active development with recent updates noted in the README.
• Community support available via Discord.
• Bug reporting and feature requests via GitHub Issues.

Licensing & Compatibility

• Licensed under the Apache 2.0 License.
• Permissive license suitable for commercial use and integration into closed-source projects.

Limitations & Caveats

Some analysis models, like Qwen2.5-0.5B-Instruct, have low JSON reliability and may require structured prompting. Other models offer moderate reliability. The effectiveness of the "insight_model" for generating structured analysis is dependent on the chosen model's capabilities.