How Ping An Life Built ChatBI: An AI‑Powered Intelligent BI Platform

Introduction

Ping An Life shares the practice and reflections of its self‑developed large‑model intelligent reporting product, ChatBI.

Outline

The discussion is organized into six parts: 1) Project background and goals, 2) Solution, 3) Product effects, 4) Landing challenges, 5) Summary and outlook, 6) Q&A.

1. Project Background and Goals

In 2023 China’s digital economy surpassed 50 billion, highlighting the contribution of innovative products like ChatBI. The product was motivated by three factors:

Traditional BI tools face technical bottlenecks in metrics and forecasting, and their user experience is sub‑optimal.

The breakthrough of GPT in text and image generation provides a solid foundation for enterprise deployment.

Enterprises increasingly prioritize digitalization and BI development.

Ping An Life therefore pursues three transformation directions: freeing the hands (automated reporting), freeing the mind (intelligent analysis), and prescribing solutions (actionable insights).

2. Solution

ChatBI can be realized in the large‑model era for four reasons:

Language capability – the model understands natural‑language syntax and semantics.

Learning ability – Retrieval‑Augmented Generation (RAG) enables rapid domain‑specific knowledge acquisition.

Tool calling – Agent orchestration allows quick invocation of existing tools and code generation.

Logical reasoning – combined model and human analysis can detect anomalies and root causes.

Product goals target a BI 3.0 era with three user demands: intelligence (model‑driven analysis suggestions), automation (auto‑generated visual reports), and real‑time response (second‑level data retrieval). The aim is to serve managers, frontline staff, and even B2C customers with comprehensive digital services.

Foundations

Complete data middle‑platform with rich data domains.

Long‑term data governance delivering tens of thousands of standardized metrics.

Reusable visualization components.

API‑enabled data services.

Private deployment and fine‑tuning capability for large models.

These strengths enable rapid construction and rollout of ChatBI.

Overall Architecture

The solution is divided into four layers:

Data middle‑platform : stores various data domains and metrics.

Platform layer : integrates API services, knowledge management, the large model, Cube/GS platforms, and a visual platform, covering question‑to‑code‑to‑visualization flows.

Agent layer : four agent types – question answering, analysis, data interpretation, and common capabilities.

Application layer : implements three core functions – What (zero‑code, real‑time query), Why (root‑cause analysis and insight), and How (prescriptive recommendations).

Technical Architecture

Core services include public services, the BI large model, data middle‑platform, and knowledge base. Built on top are five modules:

Front‑end plugins for multi‑platform access, authentication, and gateway control.

Multi‑turn dialogue leveraging contextual understanding.

Agent orchestration that acts as the system’s brain, invoking tools and knowledge bases.

AI+BI toolbox with domain‑specific small models (e.g., forecasting, anomaly detection).

Visualization system that quickly assembles charts via reusable layouts.

3. Product Effects

Key use cases demonstrate ChatBI’s capabilities:

Conversational query : Users ask natural‑language questions (e.g., agency performance). The system parses intent, retrieves data in seconds, and visualizes results.

Answer on demand : Users retrieve metric metadata and definitions instantly.

SQL generation & recommendation : The system can auto‑complete missing query parts and provide ready‑to‑run SQL or Python code for analysts.

4. Landing Challenges

Model hallucination – inconsistent answers for the same question; mitigated by knowledge‑base fallback and data‑platform verification.

Root‑cause analysis – requires extensive metric lineage and knowledge graphs, demanding significant effort to build specialized small models.

Permission management – fine‑grained row and column level controls need robust data governance.

5. Summary and Outlook

The project delivers six major values:

Broad coverage – not limited to management, but available to all employees and customers.

24/7 online availability.

Seamless integration across multiple client platforms.

Standardized, enterprise‑wide data.

Intelligent analysis that lowers the barrier to insight.

Complete insight delivery – from data to conclusions and actionable recommendations.

6. Q&A

Q1: Is ChatBI built on a dedicated large model or a fine‑tuned general model? What is the focus between Python and SQL?

A1: Ping An uses a privately deployed Qwen‑72B model, fine‑tuned and optimized for finance. SQL generation is the primary use case; Python is planned for deeper data analysis scenarios.

Q2: How is row‑ and column‑level data permission managed?

A2: A dedicated permission service checks both row and column rights against pre‑defined metric access policies.

Q3: How does the fallback strategy address model hallucination?

A3: Bad cases are collected, analyzed, and fed back into intent recognition and knowledge‑base enrichment; user‑liked answers are prioritized for further review.

Q4: What inputs are needed for effective root‑cause analysis?

A4: Metric lineage, correlation, and temporal lag information must be encoded into a knowledge graph for the model.

Q5: Does the knowledge graph include entities beyond metrics?

A5: Currently it focuses on metric relationships; other entity graphs are not part of the internal product.