Building a ‘One‑Sentence Bank’: Big Data and AI Fusion for Small Banks

1. Big Data Development History and Outlook

From the banking perspective, the core mission of big data is to drive intelligent operational decisions. Four historical goals are identified: massive physical data centralization via data warehouses, the rise and fall of thematic data marts, the pursuit of real‑time streaming data through platforms like Kafka, and the current focus on cross‑modal analysis that combines structured, unstructured data with large‑model AI capabilities.

2. Comparison of Heterogeneous Data Management Modes

Traditional data‑warehouse layering (raw, detail, summary) suffers from two shortcomings: it records only final results, missing process data, and it ignores non‑structured data and knowledge layers. A corresponding governance framework for unstructured data is proposed, mirroring the layered approach of structured data while ensuring logical continuity between software engineering and data systems.

The proposed layers for unstructured data are:

Original file layer – raw ingestion and preview without parsing.

Corpus detail layer – building a data catalog and semantic tags, e.g., labeling data security levels.

Corpus summary layer – focusing on semantics, supporting retrieval and enrichment.

Corpus application layer – leveraging general large models to construct usable knowledge.

3. Paradigm Shift from Data Engineering to Knowledge Engineering

The industry is moving from data‑centric pipelines to knowledge‑centric pipelines to better serve AI. A concrete example shows how a raw transaction record “Zhang San, 31 Dec, withdraw 5000 CNY” is transformed into a feature “Zhang San made the 10th withdrawal in the past hour”, which is then evaluated as a high‑risk indicator. This illustrates how feature engineering extracts actionable knowledge that large models can reliably consume.

4. Introducing LLMOps to Build High‑Quality Enterprise Knowledge Bases

Beyond technical challenges, management issues are addressed by extending DataOps to LLMOps. The framework covers knowledge asset operation (iteration and updates), multimodal knowledge fusion, and knowledge permission and quality control, emphasizing the banking sector’s strict data‑security requirements.

The LLMOps workflow is divided into three stages with six steps:

Knowledge acquisition – file synchronization and preview.

Knowledge assetization – catalog organization.

Knowledge cleaning – data cleansing.

Knowledge enrichment – adding QA or expert experience via templates or large‑model assistance.

Knowledge update – comparing and updating files/QA.

Knowledge retrieval – vector‑based indexing, recall, and ranking.

5. Hybrid Data System for Marketing Scenarios

In a banking marketing use case, customer behavior traces (e.g., high‑net‑worth client churn) are captured via embedded tracking. The article lists a typical decision path: product reminder → comparison of yields → investigation of transfer limits → final large‑amount transfer. Challenges include massive low‑density raw data, limited analyst bandwidth, and insufficient structured signals.

The proposed solution enriches structured data with real‑time updates (T+1) and integrates unstructured behavior vectors (e.g., multi‑channel interaction logs). Automated tagging (e.g., “investment‑active”) enables 24/7 customer analysis and personalized product recommendation.

6. Ultimate Goal: The ‘One‑Sentence Bank’

The final vision is a bank that, after a single user query, not only provides an answer but also completes the requested task. This requires moving from pure large‑model Q&A agents (e.g., Yuanbao, Doubao) to intelligent agents that combine question answering with execution capabilities.

A four‑layer architecture is described using a vacation‑request scenario:

General internet large model – basic Q&A.

Localized model with RAG knowledge base – adds internal policy answers.

Localized model + enterprise RAG + big‑data – incorporates personal HR data for personalized responses.

Localized model + big‑data + intelligent agent – recognizes intent, creates a leave request, gathers missing details, and closes the loop.

Applying the same hierarchy to a bank transfer request (“Transfer 500 CNY to Zhang San”) yields a five‑step workflow: intent recognition, tool discovery (MCP protocol), secure call, execution of business logic, and user feedback with receipt and transaction ID.

7. Q&A Highlights

Q1: Balancing rapid business response with unified data standards requires front‑loading data‑standard work and strong executive commitment.

Q2: Early signs of data‑governance failure are an over‑emphasis on speed without governance.

Q3: Designing a data‑service API can follow a data‑centric model (expose whatever data exists) or a demand‑driven model (build services based on consumer needs), with considerations for real‑time query conversion, partitioning, and portal‑style cataloging.