Tencent OlaChat: Intelligent Data Analysis Platform – Research, Architecture, and Capabilities

The Tencent PCG OlaChat team shares their research and practical experience in the field of intelligent data analysis, introducing the motivations, design principles, and key achievements of the OlaChat platform.

The DIKW model is explained as a framework that transforms raw data into information, then knowledge, and finally wisdom, highlighting the importance of extracting meaningful insights from unprocessed data.

The evolution of data analysis platforms is traced from the first generation of traditional BI (circa 1996) to agile, drag‑and‑drop BI (around 2011) and the current intelligent BI era (since 2019), where natural‑language interaction lowers the barrier for all users to become data analysts.

The rise of large language models (LLMs) is shown to bring unprecedented opportunities to intelligent data analysis, with improvements in language understanding, tool usage, logical reasoning, and rapid domain adaptation through few‑shot fine‑tuning.

Key challenges for third‑generation intelligent data products are identified: limited LLM capabilities (hallucination, lack of domain knowledge, context length), low accuracy of intelligent BI, and the need for a full‑link, end‑to‑end solution.

OlaChat’s functional overview includes a one‑stop query interface that discovers tables, generates analysis charts, and presents conclusions; a Copilot for code assistance; an IDE for SQL debugging and optimization; and the WeTable assistant for spreadsheet‑style analysis.

The OlaChat Agent design is detailed with six modules—Intention Enhance, Controller, Memory (RAG), Reasoning, Active Learning, and Voting—working together to understand user intent, retrieve structured metadata, perform multi‑step reasoning, and produce reliable answers.

The overall system architecture is divided into four layers: a unified front‑end (Copilot, Polit, and magic‑wand components), a unified back‑end service layer (APIs and orchestration), an Agent layer (composing atomic capabilities), and a common service layer (metadata retrieval, model scheduling, and annotation).

In addition to product development, the team has released several academic papers (e.g., on Text2SQL, SQL correction, and metadata‑enhanced RAG) and open‑source code repositories, demonstrating the scientific contributions behind the platform.