Technical Practices of Tencent's Intelligent BI System: Architecture, Model Fine‑Tuning, and Agent Design

With the rapid development of big data, traditional Business Intelligence (BI) products are evolving toward intelligent BI; Tencent shares its technical practice, covering engineering architecture, model fine‑tuning, guided completion, and front‑end command‑layer design to improve data analysis intelligence.

The main challenges include unclear architecture caused by fast iteration, overly complex single‑agent designs, difficulty implementing large language models (context management, long‑text memory, hallucination), and integrating traditional data analysis capabilities into intelligent scenarios.

Solutions focus on a unified architecture that adapts flexibly across products, enhancing user experience through faster, more accurate agent responses, and integrating AI capabilities via the OlaChat framework.

OlaChat is an AI‑powered data assistant that supports intent understanding, multi‑turn dialogue, query rewriting, metric discovery, data finding, SQL generation, error correction, and end‑to‑end analysis, providing a native AI data experience across various data products.

The technical stack consists of a front‑end component layer (copilot, magic wand, input modules, large‑screen analysis), an orchestration layer that uses DAGs to sequence atomic agents and services, an atomic‑agent layer that decomposes AI abilities (intent detection, Text2SQL, data interpretation) into reusable components, and a data‑service layer offering SQL services, headless‑BI metric services, and chart recommendation services. Tencent’s Mixtral large model is fine‑tuned for domain‑specific tasks.

Agent orchestration includes single‑agent multi‑step SQL correction (step 1: explain error, step 2: suggest fix, step 3: provide corrected SQL) and multi‑agent collaboration for tasks such as metric analysis, Text2SQL generation, and DSL conversion. The system handles both single‑layer SQL ( group by , having , where ) and multi‑layer SQL, converting SQL to a domain‑specific language (DSL) and then to front‑end commands.

Monitoring tracks request latency and execution details for debugging, while deployment options span SaaS, API, independent deployment, and cloud‑native solutions. Future plans include a native SQL IDE, data‑warehouse intelligence, automated annotation, canvas‑guided analysis, and continued AI capability enhancements.