AVA 4.0 Release

Getting AI to write a SQL query or generate a chart is no longer the hard part. Real business questions bring a different set of challenges: data lives in separate systems, metrics follow different definitions, and conclusions need to stand up to scrutiny.

AI-powered data analysis takes more than a stronger model. It needs trustworthy data context, reliable computation, and clear visualizations.

That is what AVA 4.0 is designed to address. AVA is an open-source data analysis and visualization framework for JavaScript applications and AI agents.

AVA 3.0 focused on automated insights and chart recommendations, using rules and statistical methods to help developers identify anomalies and trends and choose suitable charts. Version 4.0 extends that foundation to AI agents, connecting data access, analysis execution, and visualization. An agent can interpret a question using the relevant business definitions, ask AVA to query and compute the data, and present the results in a chart.

💡 Here are a few highlights of AVA 4.0:

🔗 Unified access to heterogeneous data	🔍 Analysis you can verify	🕵️ Automated insight discovery
🔮 Control over data and execution environments	📊 Intelligent chart recommendations	🔗 Easy integration with existing AI agents

If you know PandasAI, think of AVA as a data analysis and visualization framework for AI agents in the JavaScript ecosystem.

Start with a business question

A business review often begins with a question like this:

“Net sales are down. Should we increase ad spend?”

The question asks both what is happening and what to do next. Answering it means first agreeing on what net sales includes, which periods to compare, and how to attribute refunds. Then you need to find where the decline comes from and assess whether the available data supports a recommendation to spend more on advertising.

Using the same simulated retail dataset, we ask three questions in sequence:

1. Should we increase ad spend after a decline in net sales?

2. Have we calculated net sales correctly?

3. How do two departments define and calculate sales differently?

These examples draw on AVA's capabilities for understanding data, executing analysis, and visualizing results. Developers can bring those capabilities into an agent through a Skill, the CLI, or the SDK, so it can investigate business questions using data.

Establish shared business definitions

Reliable analysis depends on business definitions that people and agents can reuse: how metrics are calculated, which time periods records belong to, and how datasets relate to one another. Clear definitions provide a common basis for channel breakdowns, product analysis, and comparisons over time.

AVA defines a unified DataModel that supplies standardized context for query generation:

Large language models are powerful, but translating natural language into accurate SQL is still difficult. Using GLM-5.1, we compared performance with different levels of data context on the public DataBench analysis benchmark and the SQLite subset of the Spider2-Lite natural language-to-SQL (NL2SQL) benchmark.

On the Spider2-Lite SQLite subset, accuracy was 25.9% with only column names and sample data. Adding schema information and statistical profiles brought it to 33.3%. Including business definitions raised it to 80.7%. More context also increases token usage, of course.

Answering business questions correctly requires a semantic layer, too. The schema tells the model which fields exist; business definitions explain how to calculate metrics and relate the data. Capturing that knowledge lets people and agents work from the same definitions.

Connect heterogeneous sources for complex analysis

Orders may live in a database, advertising data in files, and operational records in a Feishu spreadsheet. A sales review often requires filtering, joining, and aggregating large volumes of records across these sources.

AVA divides the work accordingly: the model guides the investigation; code does the computation. The model breaks down the question, generates queries, and chooses the next step. Execution engines process the data and return results or errors so the model can compare findings and drill down. The full raw table does not need to go into the model's context.

AVA provides a unified interface to more than 10 data sources, so applications can reuse analysis workflows across different sources.

Once the data is connected, AVA can use different execution engines according to where the data lives and what the analysis requires:

• File analysis: DuckDB 2.0 handles filtering, aggregation, and joins across tables, supporting file queries over roughly 600 million rows.
• Database analysis: ClickHouse and Supabase run computations in the database, using existing compute resources and returning the query results.
• Code-based computation: Python/pandas handles complex transformations and numerical work, while JavaScript supports lightweight processing in a browser or Node.js.

Use visualization to guide exploration

An overall number can hide very different patterns underneath. In the sales example above, comparing channels reveals differences behind the overall decline. In other business questions, looking at distributions or comparing groups can uncover patterns that aggregates miss.

The following examples use simulated business data to show how computation and visualization work together:

Business question	Analytical approach	Visualization
Average order value hasn't changed. Should we keep the same promotion strategy?	Each period contains 2,000 orders, with an average order value of about CNY 800. Yet the share of orders below CNY 600 rises from 10.4% to 55.5%, while high-value orders also become more common. A distribution plot using kernel density estimation (KDE) reveals the shift and helps assess promotions for different customer segments.
Should all 120 stores have the same operating targets?	Monthly order volume and average order value distinguish volume-driven, balanced, and high-order-value stores, providing a basis for setting targets and allocating resources by group. The analysis uses standardization and KMeans clustering.

Choosing the right chart helps you see the issue clearly. AVA organizes chart use cases, data requirements, and limitations into heuristic knowledge. It combines that knowledge with the analytical intent and the data's characteristics to guide the model toward an appropriate visualization.

Recommendations are rendered by AntV GPT-Vis, which supports more than 26 visualization types.

Make conclusions traceable and verifiable

Business analysis often involves confirming metric definitions, comparing data, and breaking down the results at several levels. Each judgment builds on earlier findings, so checking a conclusion means following that analytical path.

AVA gives the agent the executed query, result data, and execution status for each step, making the evidence behind multi-step analysis traceable and helping pinpoint steps that need further investigation.

For example, net sales from organic search fell by CNY 877,200. You can follow the business definitions, executed query, returned data, and chart to check each assumption: Were refunds deducted? Were consignment sales counted as commission only? Did the query follow the agreed definitions? Were the display units consistent? Verifying the conclusion becomes a matter of checking specific steps in the analysis.

From verifying a sales change to investigating its causes, each result becomes evidence for the next judgment. A reasoning graph lays out that progression, helping users understand how a conclusion was reached and return to the steps that support it.

When you can inspect the process, you know where to make corrections. If a business definition is wrong, revise it. If a query is wrong, adjust the calculation. If the evidence is insufficient, keep investigating. Trustworthy analysis comes from making each step open to verification and correction.

Build faster with reusable analysis tools

Building a data analysis agent means handling data access, understanding schemas, executing queries, and generating charts. AVA packages these common capabilities as composable tools, reducing repetitive development so you can focus on business definitions, analytical questions, and the product experience.

Already using an AI agent? Install the AVA Skill with one command to let it analyze data with AVA:

npx skills add antvis/AVA

Then start a new agent session and ask a question:

Use the AVA Skill to analyze ./sales.csv (download the sample file): compare sales in January and February 2025 and identify which products contributed most to the decline.

AVA handles the queries and computations, leaving the agent to interpret the question and organize the analysis.

Building your own application or agent? Follow the SDK quickstart to combine data access, analysis, and chart generation as needed. You can build a minimal data agent with 53 lines of core code. Or use the CLI to run analyses from a terminal or script after two setup steps: installation and model configuration.

Try AVA 4.0

AVA website	Sive	GitHub
Read the docs and examples to start integrating. 🪄 Explore	A free, publicly available AI data analysis platform built on AVA. 🪄 Try it	Get the source, report issues, and contribute. 🪄 Visit

If AVA is useful to you, give it a ⭐️ on GitHub to help more developers discover it.