Tencent OlaChat: An LLM‑Powered Intelligent Business Intelligence Platform – Architecture, Capabilities, and Practice

In the rapidly evolving data analysis landscape, intelligent analytics platforms are transitioning from traditional BI to agile analysis and now to LLM‑driven intelligent BI, enabling users to interact with data via natural language and reducing learning costs.

The talk outlines four main parts: the shift from traditional BI to intelligent BI, new possibilities brought by LLMs, the practical implementation of Tencent's OlaChat platform, and a Q&A session.

Traditional BI suffers from top‑down workflows, long development cycles, and high communication overhead, while agile analysis improves accessibility but still requires significant user training.

With the rise of large language models, intelligent BI can understand user intent, generate SQL, and provide intuitive insights, turning every user into a data analyst.

The presentation reviews the development of LLMs—from early probabilistic models, through word2vec and LSTM, to Transformers, BERT, GPT, and today’s trillion‑parameter models—highlighting their impact on data intelligence.

Two metadata‑retrieval strategies are described: FlattenedRAG, which converts structured metadata into natural‑language text for conventional retrieval, and StructuredRAG, which leverages the hierarchical nature of metadata for more precise searching.

Challenges of Text2SQL in real‑world scenarios (privacy, model hallucination, data quality, noise) are addressed by a fine‑tuned LLM combined with an agent framework, selective field feeding, few‑shot prompting, and post‑generation SQL validation.

OlaChat’s core components include a multi‑turn dialogue system, task orchestration engine, AI‑BI toolbox (query rewriting, Text2SQL, metric analysis), and shared services such as unified LLM scheduling, metadata search, and annotation systems.

The architecture stacks from low‑level services to public services, agents, unified backend, and unified frontend, delivering a seamless user experience for data querying, insight generation, and analysis.

The session concludes with a Q&A covering model sizes, attribution accuracy, SQL correction and interpretation, and the trade‑offs between direct SQL generation and semantic simplification.