antvis/AVA

AVA is an AI-native visual analytics framework built for agents that automates data analysis and visualization generation.

The tool addresses the challenge of extracting insights from unstructured data by combining natural language processing with deterministic data profiling. Users can query datasets in plain English, and the system recommends analysis approaches based on data characteristics. AVA processes queries through an LLM-powered workflow that handles data loading, processing, analysis, and visualization code generation. The framework separates concerns into modular components for data handling, analysis logic, and visualization output, allowing flexibility in how analysis is performed and where it runs.

AVA suits teams building data applications that need to reduce manual analytics work and enable non-technical users to explore data through conversation. The tool runs in both Node.js environments with access to DuckDB, Python/pandas, and remote databases, and in browsers using a JavaScript interpreter engine. This dual-environment capability means you can embed it in web applications or use it server-side depending on your infrastructure. The framework's modular design means you can adopt individual components rather than committing to the entire stack.

The project maintains active development with regular updates to its analysis capabilities and agent-based workflows. Evaluation benchmarks are published and updated, showing analysis accuracy across different workflow configurations and LLM models. The codebase includes comprehensive documentation of evaluation methodology, timing metrics, and token usage for different analysis approaches, enabling informed decisions about which workflow configuration suits specific use cases.