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 Business Intelligence (BI) to agile and finally to AI‑driven intelligent BI, with large language models (LLMs) enabling natural‑language interaction and reducing user learning costs.

The presentation is organized into three main parts: the shift from traditional BI to intelligent BI, new possibilities introduced by the LLM era, and the practical deployment of Tencent's OlaChat platform.

Traditional BI suffers from a top‑down workflow, long development cycles, and high communication overhead, often requiring users to wait days for analysis results.

Agile analysis, spurred by the mobile‑internet boom, allows self‑service data exploration through drag‑and‑drop interfaces, yet still presents a steep learning curve for complex operations such as period‑over‑period calculations.

Early intelligent BI concepts emerged around 2019, aiming to turn every user into a data analyst by simplifying the analysis pipeline and lowering technical barriers.

The evolution of LLMs—from probabilistic models, through neural networks (word2vec, LSTM), to Transformer‑based models (BERT, GPT) and trillion‑parameter systems—has dramatically improved language understanding, generation, and logical reasoning, opening new opportunities for data analytics.

LLMs enhance intelligent BI in four key ways: superior language ability for intuitive query interpretation, tool usage that converts natural language into API calls or code, improved logical reasoning for pattern and trend detection, and in‑context learning that reduces the need for extensive fine‑tuning.

OlaChat, built on Tencent PCG’s “Ola” asset‑management platform and the “Lantern” analytics engine, leverages rich metadata and user‑behavior logs to provide NL2SQL, audience insights, and other analytics services, dramatically lowering the barriers to data retrieval and usage.

The platform’s core modules include a multi‑task dialogue system, a task orchestration engine, an AI + BI toolbox (query rewriting, Text2SQL, metric analysis, etc.), and common services such as unified LLM scheduling, knowledge‑retrieval enhancement, and a annotation system for domain‑specific understanding.

The dialogue system offers context understanding, intent recognition, natural‑language understanding (NLU), dialogue state tracking (DST), dialogue policy (DPL), and natural‑language generation (NLG) to interact with users seamlessly.

Metadata retrieval faces challenges due to the structured nature of tables and metrics; two RAG strategies are proposed: FlattenedRAG, which converts structured metadata into natural‑language passages for standard semantic search, and StructuredRAG, which first retrieves core elements (metrics, dimensions) and then performs a second‑stage search to refine results.

Text2SQL encounters issues such as data privacy, model hallucination, low robustness, and scarce high‑quality training data. OlaChat addresses these by fine‑tuning a 70B LLM with an agent framework, generating synthetic data through collection, anonymization, random sampling, and augmentation, and enforcing accuracy and diversity checks across difficulty levels.

Post‑generation, SQL statements are validated for executability and semantic alignment, with a correction loop that leverages LLMs and auxiliary information (metadata, error messages) to improve reliability; active learning further focuses on frequent failure patterns.

Beyond Text2SQL, additional agents support query rewriting, error correction, optimization, interpretation, and Q&A, providing a comprehensive intelligent‑analysis ecosystem.

The overall architecture stacks low‑level services, shared public services, agents, unified back‑end and front‑end, delivering a cohesive user experience across various analytics scenarios.

The Q&A session reveals that a lightweight 8B model handles quick data fetching, while a 70B model is fine‑tuned for NL2SQL; accuracy relies on combining LLM reasoning with external attribution tools and metadata to ensure reliable results.