VizGen: AI-Driven Graph Generation

• VizGen is an AI-assisted graph generation and data exploration system that converts natural-language requests into SQL-driven visual analytics.
• It integrates multiple agents for intent classification, SQL generation, visualization, and explanation to automate end-to-end data analysis.
• By enabling conversational refinement and context-rich insights, VizGen minimizes technical barriers and enhances interactive data exploration.

Searching arXiv for the primary VizGen paper and closely related visualization systems to ground the article in current literature. Searching arXiv for "VizGen" and related genomics visualization systems. VizGen is an AI-assisted graph generation and data exploration system that converts natural-language requests into database-backed visualizations, analytical findings, and explanatory responses. It is presented as a conversational, multi-agent analytics system rather than a generic plotting library: the system classifies user intent, generates and validates SQL, recommends graph types, renders visualizations, supports conversational customization, extracts patterns and anomalies, and augments explanations with contextual information gathered from the internet. Its stated objective is to reduce the technical burden of querying databases, selecting chart types, customizing graphics, and interpreting results, thereby broadening access to visualization-driven analysis for non-technical and mixed-expertise users (Fernando et al., 26 Sep 2025).

1. Definition and problem setting

VizGen is framed around a specific usability barrier in contemporary data analysis. Conventional tools such as Microsoft Excel, Tableau, and Power BI are described as powerful, but still often require knowledge of data structuring, data cleaning, chart selection, visual encoding, database querying, and result interpretation. VizGen addresses that barrier by allowing requests in plain language, after which the system handles natural-language understanding, SQL generation, data retrieval, chart recommendation, graph rendering, customization, insight generation, and explanatory narration (Fernando et al., 26 Sep 2025).

The system is explicitly intended for both non-technical users and mixed-expertise users, including students, developers, data analysts, and business users. This scope is reflected in the reported evaluation population, which ranged from beginners to experts in data visualization. The paper therefore places VizGen within the class of natural-language interfaces for data analysis, but broadens that class by coupling natural-language query handling to downstream visualization and explanation modules rather than limiting the system to query translation alone (Fernando et al., 26 Sep 2025).

A common misconception is to treat VizGen as merely a text-to-chart frontend. The described system is broader: it supports insight requests such as anomaly detection, explanation requests such as causal-style contextualization of an observed trend, customization requests that refine an existing chart, and system operations such as connecting to new databases or exporting results. In that sense, the graph is one artifact in a larger routed workflow rather than the sole product of the system (Fernando et al., 26 Sep 2025).

2. End-to-end analytical workflow

The end-to-end pipeline begins with a natural-language request and ends with a unified response that may include a visualization, extracted insights, explanatory context, or a combination of these outputs. The first stage is intent classification. The paper states that the Intent Classifier recognizes intents corresponding to visualization, insight, explanation, customization, system, and other; the workflow figure also mentions schema and exploratory intents. Routing decisions at this stage determine which downstream agents are activated (Fernando et al., 26 Sep 2025).

For requests that require data access, VizGen invokes the SQL Agent. This agent comprises a Metadata Retriever, which acquires schema information from the connected database; a SQL Generator, which produces a query from the user request and schema metadata; a SQL Validator, which checks correctness and efficiency; and a SQL Executor, which runs the query on the live SQL database. Retrieved results are then passed to a Data Preprocessor, which organizes numeric, categorical, and temporal data, extracts features, and calculates cardinalities (Fernando et al., 26 Sep 2025).

Visualization requests are routed to the Visualization Agent, particularly the Graph Ranker, which ranks suitable graph types according to data characteristics, data types and structure, user preferences, and visualization best practices. If analytical intent is present, the Analysis Agent activates an Insight Generator that identifies trends, anomalies, patterns, and correlations. If explanation is requested, the Explanation Agent generates a search plan, executes search over external sources, and synthesizes those results with internally derived data insights. The final Response Generator integrates outputs from all invoked agents into a single response (Fernando et al., 26 Sep 2025).

The workflow is not restricted to single-intent interactions. The paper explicitly states that multi-intent queries activate multiple agents sequentially. A request such as “Show me a bar chart of sales and explain the biggest trend” routes through the SQL Agent, Visualization Agent, Analysis Agent, and Explanation Agent. This indicates that VizGen is architected as a routed composition of specialized stages rather than as a monolithic model call (Fernando et al., 26 Sep 2025).

3. Multi-agent architecture and implementation

VizGen’s central architectural claim is its multi-agent organization. The system is implemented with LangChain and LangGraph and is described as a state-driven graph that dynamically routes requests among specialized nodes. The principal modules are the Intent Classifier, SQL Agent, Visualization Agent, Analysis Agent, Explanation Agent, Customizer node, System node, and Response Generator. Intermediate products are passed between components, including schema metadata, generated SQL, query results, preprocessed data features, analytical findings, and external search results (Fernando et al., 26 Sep 2025).

The SQL Agent is foundational because both visualization and analysis depend on live data retrieval. The Visualization Agent includes the Data Preprocessor and Graph Ranker. The Analysis Agent centers on the Insight Generator. The Explanation Agent contains an Insight Explanation Query Generator, a Search Execution Engine, and an Explanation Generator. The Customizer node handles natural-language refinement of existing visualizations, and the System node handles operational requests such as database connection, authentication, and export (Fernando et al., 26 Sep 2025).

The implementation stack is comparatively concrete for a systems paper. The backend uses FastAPI; MongoDB stores user interaction logs, visualization metadata, preferences, and feedback; SQLAlchemy handles external database connectivity; the frontend uses React.js, Framer Motion, and Material-UI; Plotly.js renders graphs; LangChain and LangGraph orchestrate the multi-agent workflow; LangSmith supports observability and debugging; Docker is used for containerization and cloud deployment; and testing uses PyTest, Jest, and manual testing. External relational database support is reported for MySQL, PostgreSQL, MariaDB, MS SQL, and Oracle DB (Fernando et al., 26 Sep 2025).

The paper explicitly names Claude 3.7 Sonnet and Gemini 2.0 Flash as the principal LLMs used for natural-language understanding, context interpretation, SQL generation, and graph-type recommendation. It does not specify subtask allocation between these models, nor does it disclose prompt templates, confidence thresholds, or a detailed ambiguity-resolution strategy. This absence is significant because the system’s practical behavior depends heavily on such orchestration details, yet those details remain high-level in the published account (Fernando et al., 26 Sep 2025).

4. Visualization generation, customization, and explanation

VizGen’s visualization layer is centered on recommendation rather than direct specification. The Graph Ranker recommends graph types using data characteristics, user preferences, and best practices. The paper does not provide a formal scoring function, training objective, or explicit rule base for this ranking stage, but it does state that the mechanism is an LLM-based recommendation engine informed by data-feature inspection from preprocessing. Supported or intended graph forms include time-series graphs, scatter plots, histograms, bar or column charts, line charts, and heatmaps, while Plotly.js renders “various types of graphs” on the frontend (Fernando et al., 26 Sep 2025).

Customization is a first-class operation rather than a post hoc editing convenience. The paper gives the example “Change the color of this chart to blue” and states that natural-language customization can also adjust chart types and modify visual elements. Frontend controls further include filtering, zooming, and exporting. The interaction model is therefore iterative and conversational: an initial chart can be produced from a natural-language request and then refined through follow-up commands without requiring reconstruction of the visualization from scratch (Fernando et al., 26 Sep 2025).

VizGen also departs from conventional natural-language visualization systems by coupling chart generation to automated analysis. The Analysis Agent identifies patterns, anomalies, trends, and correlations through autonomous tool selection, statistical analysis, and pattern recognition. The paper does not specify the exact statistical procedures, anomaly thresholds, correlation measures, or trend estimation algorithms. As a result, the analytical layer is described systemically rather than mathematically. The same is true of insight quality: the system is designed to produce analytical findings, but the paper does not benchmark the correctness of those findings against formal baselines (Fernando et al., 26 Sep 2025).

The Explanation Agent extends this analytical layer by adding contextual information gathered from external sources. Its purpose is not to generate the chart itself, but to help answer why a pattern may have occurred. This internet-enriched explanation capability is distinctive, but it introduces clear reliability questions. The paper does not specify safeguards for source credibility, misinformation filtering, citation handling, or trust scoring. A plausible implication is that explanation quality depends not only on internal data analysis but also on the stability and verifiability of external search results (Fernando et al., 26 Sep 2025).

5. Evaluation, reported findings, and limitations

The reported evaluation is survey-based. VizGen was assessed through a Google Forms survey containing Likert-scale ratings from 1 to 5 and open-ended questions. There were 19 participants from varied backgrounds, including data analysts, developers, and undergraduate students. Commonly used prior tools included Microsoft Excel at 74%, Python visualization libraries at 53%, and Power BI at 21% (Fernando et al., 26 Sep 2025).

The paper reports average ratings of 4.1 for Clarity of Demo, 4.4 for Ease of Use, 4.4 for Confidence Without Support, 4.4 for Overall UI Design, 3.3 for Navigation Ease, and 3.9 for Speed and Responsiveness. Participant-valued features were Data summary & insight at 79%, Auto chart recommendation at 74%, Text-to-Graph generation at 68%, AI chatbot support at 68%, and Visual customization at 53%. All participants agreed that VizGen addresses a real need; 79% were willing to participate in additional testing; 95% would recommend VizGen to others; and average likelihood of future use was 4.4 out of 5 (Fernando et al., 26 Sep 2025).

These results indicate positive user perception, but the paper is explicit about what remains unmeasured. It does not provide benchmark datasets, controlled task-completion studies, NL2SQL accuracy metrics, chart recommendation accuracy, latency benchmarks, baseline comparisons against systems such as Tableau, Power BI, or NL4DV, ablation studies on individual agents, or quantitative measures of insight quality. Consequently, the evidence base is primarily user-perception-based rather than a rigorous performance benchmark (Fernando et al., 26 Sep 2025).

The limitations section is correspondingly important. The paper notes inconsistency in the initial graph recommendation system, latency introduced by more accurate text-to-SQL models, integration complexity across heterogeneous external databases, and edge-case difficulty in natural-language customization. It also leaves open risks of incorrect SQL, semantically wrong joins or aggregations, unsupported explanatory narratives, unreliable web-derived context, and under-specified privacy and security controls. These are not incidental implementation problems; they are structural risks for any natural-language, multi-agent analytics system that interleaves query synthesis, statistical interpretation, and external-context retrieval (Fernando et al., 26 Sep 2025).

6. Position within the broader visualization-systems landscape

VizGen belongs to a broader family of systems that combine visualization with layered interaction, but its problem setting differs from several adjacent arXiv systems. “GeneVis” is a web-based genetics visualization tool that integrates gene clusters, gene–gene interactions, and gene–disease associations through coordinated cluster and gene views; its central design pattern is overview-to-detail exploration over heterogeneous biological datasets (Leeuwen, 2018). “Gosling Designer” is an end-to-end platform for creating, exploring, managing, and sharing genomics visualizations through the Gosling grammar, workspace-based data management, and role-based collaboration (L'Yi et al., 28 Oct 2025). “VizGenie” targets scientific visualization of complex volumetric datasets through an agentic framework that can generate, validate, cache, and reuse VTK-based visualization modules, with vision-language support for feature-centric requests such as “visualize the skull” (Biswas et al., 18 Jul 2025).

This suggests that VizGen’s distinctive contribution lies not in genomics specificity, ontology visualization, or scientific-volume rendering, but in natural-language mediation over SQL-backed, general-purpose visual analytics. GeneVis demonstrates coordinated multiscale biological exploration; Gosling Designer demonstrates declarative genomics authoring and collaboration; VizGenie demonstrates self-refining tool generation for scientific visualization. VizGen instead emphasizes natural-language intent routing, NL-to-SQL execution, graph recommendation, conversational refinement, automated insight extraction, and context-enriched explanation in a single web application (Fernando et al., 26 Sep 2025).

A second point of contrast concerns canonical artifacts. In Gosling Designer, the portable artifact is the declarative Gosling specification (L'Yi et al., 28 Oct 2025). In VizGenie, successful generated visualization modules are stored and reused via a persistent cache (Biswas et al., 18 Jul 2025). In VizGen, the primary artifact is the generated response assembled from SQL results, Plotly-rendered charts, analytical findings, and optionally external-context explanations, while MongoDB stores interaction logs, metadata, preferences, and feedback rather than a domain-specific declarative grammar (Fernando et al., 26 Sep 2025). This difference matters because it shapes reproducibility: VizGen is reproducible at the workflow and logging level, but the paper does not define a portable visualization-specification layer analogous to Gosling.

The broader significance of VizGen is therefore architectural. It treats data visualization as a routed analytics workflow comprising intent recognition, schema-aware query generation, chart recommendation, iterative customization, automated analysis, and explanatory augmentation. Its limitations are equally architectural: absent benchmarked NL2SQL accuracy, unformalized chart ranking, unspecified analytical methods, and limited safeguards for externally sourced explanations. Within that balance, VizGen exemplifies a 2025 tendency to reconceive visualization systems as agentic, conversational, database-coupled analytical environments rather than as standalone charting tools (Fernando et al., 26 Sep 2025).