How Chat BI Transforms Data Warehousing with AI: Unlock Real‑Time Insights

Introduction

The session, led by iQIYI Technical Director Zhang Xiaoming, explores the intelligent upgrade of iQIYI’s data middle‑platform and BI system, focusing on the lake‑warehouse integration and the fusion of data warehouses with Chat BI.

Eight Core Modules

Chat BI background introduction

Chat BI core implementation process (three‑step method + detailed decomposition)

Chat BI technical architecture design (layered decomposition + precise recall)

Key challenges and solutions in building Chat BI

Practical experience for improving Chat BI effectiveness

Core process of data‑warehouse integration

Data insight based on RAG knowledge base

User operations: cold‑start and promotion strategies

BI Evolution

BI systems have progressed through three versions:

BI 1.0 (Report era) : Serves business leaders; long development cycles; users retrieve data via SQL, resulting in low efficiency.

BI 2.0 (Agile era) : Enables self‑service with drag‑and‑drop; leverages OLAP for faster queries; developers must model data and publish datasets.

BI 3.0 (Chat BI era) : Adds large language models on top of BI 2.0; users interact via natural‑language Q&A, lowering the learning curve and improving data accessibility.

Chat BI Core Implementation

The macro workflow consists of intent understanding, data development, and testing, each further decomposed into detailed steps:

User intent parsing and ambiguity removal

Generating structured detailed requirements

Fine‑grained decomposition and multi‑round model verification

Data development transforms structured requirements into executable SQL, addressing hallucination and usability issues through semantic‑SQL generation, SQL calibration, and physical‑SQL generation.

Testing covers SQL usability, result correctness, and user readability, with fallback strategies for execution failures.

Technical Architecture Design

The architecture centers on metadata management, enabling precise recall and demand confirmation. It includes:

Metadata layer : Classifies ordinary and personalized metadata, stored in high‑performance retrieval engines (HLP) and RAG knowledge bases.

Demand confirmation layer : Selects appropriate LLMs (e.g., Qwen, DeepSeek) to generate accurate SQL.

SQL generation layer : Multi‑model collaboration with syntax checking.

Execution and feedback layer : Uses StarRocks for OLAP queries, Spark for large‑scale processing, and Redis for caching, forming a query‑execution‑feedback loop.

Data‑Warehouse Integration

Prioritizes wide tables to support diverse user queries, aligns metadata with user perspectives, and follows a three‑step process: theme matching, metric filtering, and dimension filtering to retrieve relevant datasets.

Standardization procedures differ for already modeled assets (direct adaptation) and unmodeled assets (modeling before integration), ensuring consistent data structures and metric logic.

Data Insight with RAG

RAG knowledge bases combine static documents and dynamic APIs to provide deep analysis, supporting data interpretation and attribution across multiple data sources.

User Operations: Cold‑Start and Promotion

Target high‑demand users (analysts, product managers, operators) with low entry barriers, showcase high‑frequency pain‑point queries, define a north‑star metric (SQL generation accuracy), continuously monitor Q&A logs, and close the feedback loop to improve trust and adoption.

Conclusion

The talk summarizes the end‑to‑end workflow of Chat BI, from architectural design and data‑warehouse integration to performance optimization and user‑centric operation, demonstrating how AI‑driven BI can unlock data value at scale.