Summary

Predikit bridges the gap between machine learning models and AI agents, enabling developers to transform trained scikit-learn or XGBoost models into LLM-callable tools. It automates the generation of JSON schemas, enforces typed I/O, and significantly reduces boilerplate code, making ML model integration with AI agents seamless and efficient.

How It Works

The core of Predikit lies in its ModelTool class, which wraps compatible ML estimators. It automatically generates OpenAPI-compliant function-calling schemas and Pydantic input schemas, allowing LLMs to understand and invoke the underlying model. ModelEnsemble allows for parallel execution and reconciliation of multiple models using strategies like averaging or voting, while ToolRegistry groups tools for bulk export to LLM frameworks.

Quick Start & Requirements

• Install: pip install predikit
• XGBoost Support: pip install predikit[xgboost]
• LangChain Support: pip install predikit[langchain]
• MLflow/Snowflake Support: pip install predikit[mlflow] or predikit[snowflake]
• Prerequisites: Python, fitted scikit-learn or XGBoost models.
• Resources: A 30-second example is provided in the README demonstrating basic usage.

Highlighted Details

• ModelTool & ModelEnsemble: Wraps individual models or ensembles into LLM-friendly tools with typed inputs/outputs and auto-generated schemas.
• LLM Framework Integration: Directly exports tools as OpenAI function schemas (.to_openai()) or LangChain tools (.to_langchain()).
• Strict Input Validation: Leverages Pydantic for robust input validation, enforcing that schema field names precisely match model training feature names.
• Confidence Routing: Supports classifiers with predict_proba to handle low-confidence predictions via warn, raise, or fallback modes.
• Model Registry Loaders: Built-in support for loading models directly from MLflow and Snowflake Model Registries.

Maintenance & Community

The project includes a CONTRIBUTING.md file detailing development setup and guidelines, and a CHANGELOG tracks notable changes. Specific community channels or active maintainer information are not detailed in the README.

Licensing & Compatibility

The project is released under the permissive MIT license. This license allows for broad compatibility with commercial and closed-source applications, with minimal restrictions.

Limitations & Caveats

Currently, predikit primarily supports scikit-learn and XGBoost models. Support for HuggingFace, PyTorch, and TensorFlow models is planned for future releases. The strict field naming rule between Pydantic schemas and model features requires careful attention to avoid runtime ValueError exceptions.