Data Studio (AI agent)

The Data Studio is woku’s conversational analytics module. Instead of configuring a report field by field, you describe what you need in natural language and an AI agent builds an insights report on your company’s real data: KPIs, evidence-backed findings, visualizations, and recommended actions, ready to review and publish. It complements the report builder: the builder is deterministic and schedulable (ideal for repeatable deliveries in CSV/Excel/PDF); Data Studio is exploratory and conversational (ideal for getting tailored insights in minutes).

​

How it is used

1. You open a conversation and describe the report you want.
2. The agent proposes a plan (what data it will use and what it will show) and waits for your approval.
3. Once you approve it, it queries the data on the server, analyzes it, verifies every number against the datasets, and builds the report.
4. You can request changes in natural language (including changing the analyzed period), go back to a previous version, or compute advanced metrics such as predictions.
5. The report is versioned; you can publish it with a public link or one protected by a key.

Throughout the whole process the conversation receives live events (the agent “thinking”, the plan, the extracted data, the visual review, and the finished report).

​

How the agent works

The agent is a conversational orchestrator that proposes a plan, waits for your approval, and then runs a deterministic pipeline: it queries the data on the server, analyzes it, verifies the evidence, renders the report, and reviews it visually before delivering it.

​

The five tools

​

The report pipeline

create_insight_report is not a free agent, but a deterministic pipeline:

1. Server data: the datasets are always queried on the server side through the query catalog; the model never touches the database.
2. Analysis: the analyst produces a structured report with KPIs, evidence-backed findings, visualizations, and recommended actions.
3. Evidence verification: every numeric claim is verified against the datasets, with a corrective retry. Findings that do not pass verification are discarded: a report never publishes unverified numbers.
4. Deterministic render: the HTML is generated self-contained, with SVG rendered on the server and with no CDN dependencies.
5. See and repair: a multimodal model reviews the rendered result before delivering it, with a maximum of 2 repair rounds.
6. Persistence: the report is saved as a new version of the app.

​

Data access (multi-tenant)

The agent never queries the database directly: it works over a query catalog on the server. The companyId is always injected from the authentication context and never from whatever the model says, guaranteeing isolation between companies. The catalog covers global KPIs, metrics by woku and by folder, rating trends, clients and segments, NPS and its tools, forms and response channels.

​

Streaming events (SSE)

The agent endpoint responds as text/event-stream. These are the events it emits:

​

Persistence and publishing

• Conversation: message history, captured plan, and executed tools.
• App: metadata and immutable versions, each one keeps its own HTML. Reviewing or going back to a previous version never alters the others.
• Publishing: each app can be published and unpublished; the link uses a slug and access can be public or protected by key, with key verification when opening it.

​

Limits per conversation

To control cost and abuse, each conversation has caps:

​

Models

• claude-sonnet-5, the orchestrator conversation and the report analysis.
• o4-mini, only the multimodal visual review.
• gpt-5.5, fallback for the plan and the analysis under a transient overload.

​

Data Studio vs. report builder

Both read the same data and coexist: use the agent to explore and get insights, and the reports for governed and scheduled deliveries.