Next‑Generation Data Analysis Platform: Integrating Chat BI and Headless BI

The data analysis platform is a crucial vehicle for monetizing internal data, yet enterprises face high entry barriers, inconsistent metrics, and slow response to requests, limiting its value.

Three traditional analysis modes are identified: (1) SQL‑driven exploration, which requires high SQL proficiency; (2) drag‑and‑drop dashboards, which are limited in scope and still demand manual data interpretation; (3) static dashboards, which lack flexibility for new dimensions.

Business teams desire a self‑service model where they can directly query and visualize data, while data teams struggle with redundant data sets, inconsistent metric definitions, and permission management across layers such as ODS, DWD, DWS, and ADS.

To address these pain points, the architecture evolves by introducing Headless BI, comprising a semantic model and a semantic layer. This unifies metric definitions, enforces consistent permissions, and abstracts data access through a language called S2SQL, which can be queried via JDBC or HTTP.

Chat BI is then layered on top to improve usability, allowing users to ask natural‑language questions on PC or mobile devices, with the system generating SQL via large‑language‑model inference. Challenges such as schema linking, join complexity, and hallucinations are mitigated by integrating the semantic layer.

Key components include:

Mapper modules (Schema, Keyword, QueryFilter) that retrieve relevant semantic entities for the LLM.

Corrector modules (Schema, Grammar, Time) that fix hallucinations in generated queries.

Semantic Corrector to resolve schema, syntax, and temporal errors.

Chat Memory (short‑term and long‑term) for continual learning of domain knowledge.

Agent framework with Planner, Text2SQL, and SPI‑based extensible plugins to handle complex data requests.

The system enables business users to obtain instant answers, view definitions, and receive data summaries, while data teams focus on modeling and plugin development.

Future work will enhance data lineage, materialized view acceleration, intelligent modeling via LLMs, and richer multimodal interaction (voice, mobile) to further lower the barrier for data‑driven decision making.