Sugar BI: AI‑Driven Next‑Generation Business Intelligence Platform

This article is a comprehensive recap of the "AI‑Driven New Generation BI Platform – How to Enable Smart Decision‑Making" session from the Baidu Create Conference on April 16, 2024.

It first introduces the evolution of Sugar BI (originally Baidu’s internal reporting platform ShowX) from its 2016 launch, through a full rewrite and rebranding in 2018, to the release of versions 2.0 and 3.0 that expanded visual components and data‑source support. Since 2021 the product has pursued intelligent analysis, integrating AI capabilities to assist users in data exploration.

The platform now supports more than 30 data sources—including relational databases (MySQL, SQL Server, Oracle, PostgreSQL, domestic databases such as 达梦, 人大金仓, 华为 GaussDB), big‑data engines (Hive, Spark, Doris, ClickHouse, Presto), NoSQL and time‑series stores (InfluxDB, Prometheus, Redis, MongoDB), as well as Excel/CSV uploads and static JSON demos. An internal tunnel client enables secure access to on‑premise databases without exposing public IPs.

Sugar BI is built on Baidu’s open‑source ECharts and D3 libraries, offering over 150 visual components, 60+ screen templates for various industries, and both 2D and 3D chart types. The UI is zero‑code: users drag‑and‑drop data fields to create charts, configure interactions (filters, drill‑down, cross‑chart linking), and the layout automatically adapts to mobile devices.

From a BI perspective, Sugar BI provides data‑source connectivity, semantic data‑modeling (automatic dimension/measure classification, virtual calculated fields, multi‑table joins with nearest‑common‑ancestor optimization), a powerful BI engine (filtering, drill‑down, cross‑filtering, SQL‑level calculated fields), and a secondary calculation engine for advanced analytics such as retention, cross‑pivot, and alerting. It also leverages Presto for federated queries across heterogeneous sources.

The AI side is embodied in the Sugar Bot “intelligent Q&A” module. Powered by Baidu’s Wenxin large‑model series via the Qianfan platform, Sugar Bot converts natural‑language queries into NL2JSON (a structured representation of SQL), applies prompt‑engineering, performs hallucination correction, and finally generates the appropriate visual output using the intelligent chart recommendation engine.

Intelligent chart recommendation works by abstracting each chart type into a set of features (required dimensions/measures, mandatory bindings, optional scoring rules). When a user drags fields onto a canvas, the system extracts these features, matches them against chart‑type profiles, and ranks candidates using a scoring formula: mandatory rules are multiplied (zero if any fail) and optional rules are summed, ensuring only valid charts receive a non‑zero score.

Prompt engineering for Sugar Bot includes six parts: field schema, example data, output format, precautions (private domain knowledge, current time), few‑shot examples, and the user query. Few‑shot examples are generated either via a cold‑start rule‑based approach (common queries like “sales by city”, “last month’s profit”) or by leveraging user‑feedback vectors (liked queries) to provide more relevant context and to fine‑tune the LLM.

The overall architecture of Sugar Bot consists of four stages: prompt construction, NL2JSON generation, intervention/correction, and intelligent chart rendering. The system also supports function‑call‑based chart type specification, allowing users to enforce a particular visualization (e.g., map for geographic fields).

In summary, Sugar BI combines a robust, multi‑source BI engine with cutting‑edge large‑model AI to deliver automatic data modeling, natural‑language analytics, and context‑aware visual recommendations, positioning it as a leading AI‑augmented BI solution.