Chat With Your Data

Once the data is connected, unified, and organized within the Knowledge Layer, Numbers Station enables direct, conversational interaction. Users enter questions in everyday language rather than crafting SQL queries or searching through dashboards. Under the hood, this interaction is managed by a coordinated set of specialized software agents, each responsible for a specific function in the query resolution process. Together, these agents form a dynamic, multi-step workflow that retrieves or generates accurate, context-aware responses and can even recommend or perform follow-up actions.

Planner Agent

The Planner Agent is the central coordinator. When a user asks a question, this agent decomposes the request into a sequence of subtasks and delegates them to the appropriate specialized agents. It uses the Knowledge Layer to understand relevant entities (datasets, dashboards, metrics) and decides which agents need to be involved based on the query's complexity and required outputs. The Planner Agent handles both parallel and sequential executions, ensuring that all necessary steps—searching for existing answers, clarifying intent, generating new queries, creating visualizations, or proposing actions—are completed in the correct order.

Search Agent

The Search Agent checks whether the requested information has already been addressed by existing dashboards, previously answered queries, or stored documentation. This involves scanning indexes in the Knowledge Layer, referencing historical Q&A logs, and searching through dashboard metrics or analytical artifacts. If a complete or partial answer is found, the Search Agent returns this information, allowing the system to respond rapidly to common requests without re-computing results. This reduces unnecessary query execution and leverages the organization's existing knowledge assets.

Intent Agent

In cases where the user's question is ambiguous, insufficiently detailed, or unclear, the Intent Agent requests clarification. It uses natural language understanding techniques to detect missing context (e.g., which time period to consider or which specific metric is needed) and then follows up with the user to refine the question. By ensuring the user's intent is clearly defined, the Intent Agent improves the accuracy of subsequent steps. Once the user's needs are fully understood, it passes the clarified intent back to the Planner Agent.

Query Agent

When the question cannot be answered from existing resources, the Query Agent generates new SQL queries against connected data warehouses (Redshift, BigQuery, Snowflake, Databricks) using the Knowledge Layer to understand schemas, joins, and business logic. It iterates on the SQL it produces, reviewing query results and adjusting the logic until the returned data matches the user's requirements. This iterative process may involve adjusting filters, aggregations, or joins based on the feedback loops built into the system. By leveraging the semantic context from the Knowledge Layer, the Query Agent ensures that the final query is aligned with established definitions and data structures, reducing the risk of returning irrelevant or incorrect results.

Charting Agent

Returning raw numeric results or textual responses is not always sufficient. The Charting Agent addresses this by automatically producing visual representations—charts, tables, dashboards—tailored to the query's outcome. Leveraging metadata from the Knowledge Layer and visual AI capabilities, this agent selects appropriate chart types (e.g., line graphs for time series, bar charts for categorical comparisons) and configures them to highlight important trends or outliers. The Charting Agent ensures that end users can quickly interpret the returned data, reducing the effort needed to derive insights from raw query outputs.

Action Agent

Beyond simply answering a question, the Action Agent identifies potential follow-up actions. For instance, if the user requests a recurring report, the Action Agent can schedule it automatically. If the user needs to present findings, the Action Agent can generate a slide deck that includes charts, captions, and key metrics. By integrating with APIs and external tools (via OpenAPI/Swagger specifications), the Action Agent can fetch additional benchmarks, trigger alerts, or perform updates in downstream systems. This closes the loop from insight to execution, enabling users to move from data retrieval to tangible outcomes within the same interface.

Summary

All these agents work in concert through the Planner Agent's orchestration. As each agent completes its assigned task, it updates the Planner Agent with intermediate results. The Planner Agent then decides whether further steps are needed—such as requesting additional clarification or refining the SQL query—and passes control to the next agent accordingly. Throughout this process, the agents consult and update the Knowledge Layer, ensuring consistency, contextual understanding, and alignment with the organization's defined semantics.

The user experiences a direct, conversation-based workflow. Instead of switching between multiple tools or manually piecing together queries and charts, the user can simply ask questions, refine them as needed, and receive structured responses, visualizations, and recommended actions in one integrated environment. This approach shortens the path from question to actionable insight, helping teams make informed decisions without leaving the Numbers Station enabled interface.