ChatBI: NetEase’s AI‑Powered Business Intelligence Platform – Architecture, Capabilities, and Real‑World Applications

In the era of digital transformation, large language models have opened new opportunities for data analysis. NetEase’s ChatBI combines AI and traditional BI to provide an intelligent, low‑cost, and high‑efficiency analytics platform for enterprises.

The platform’s workflow includes data ingestion (supporting over 40 data sources), data preparation (custom SQL or drag‑and‑drop ETL), data modeling (materialized views and relational models), data application (reports, dashboards, and the ChatBI module), and data distribution (email, links, mini‑programs, mobile).

AI brings three main benefits to analytics: lower entry barriers, higher efficiency, and smarter insights. However, challenges such as hallucination, query performance, and scenario‑specific adaptation remain.

ChatBI’s architecture consists of a front‑end layer, intent recognition, a core query pipeline (pre‑processing, a self‑developed NL2SQL model, post‑processing, and a data‑query infrastructure), and a visual query engine that generates executable SQL for the underlying MPP engine.

The NL2SQL model is fine‑tuned on diverse domain data, outperforms GPT‑4 with an 84% accuracy on a 1,000‑case benchmark, and supports custom functions, UDFs, and DSL conversion for traceable and intervene‑able queries.

Performance is ensured through materialized views that transform complex joins into single‑table queries, leveraging ClickHouse’s MPP capabilities to achieve sub‑second response times even on massive datasets.

Open capabilities include single sign‑on, token authentication, data‑model synchronization APIs, plug‑in support for third‑party LLMs and RAG systems, and embeddable APIs for seamless integration into other products.

Real‑world deployments span HR chatbots, self‑service analytics for NetEase Cloud Music, tobacco industry dashboards, and university data assistants, demonstrating significant improvements in query speed, user adoption, and data‑driven decision making.