Project Background
As a top-tier insurance company in China, Pacific Insurance has a vast collection of business documents including insurance clauses, claims rules, product materials, etc., totaling over 30,000 documents. Claims adjusters, sales consultants, and new employees need to consult these documents frequently every day. Traditional keyword search struggled to precisely locate information; employees spent an average of 15 minutes to find the needed content, with accuracy below 40%.
Core Pain Points
Solution
Private RAG Architecture Deployment
A RAG (Retrieval-Augmented Generation) system based on Qwen2.5-72B was deployed within the internal network environment. Over 30,000 documents underwent structured chunking and vector indexing to build an enterprise-grade knowledge graph. All data processing and inference were performed within the internal network, meeting the financial industry's compliance requirement of keeping data within the domain.
Intelligent Q&A and Knowledge Recommendation
A natural language Q&A interface was implemented, supporting multi-turn conversations and contextual understanding. The system not only returns precise answers but also provides original clause citations and related knowledge recommendations, helping users fully understand business rules.
Automatic Knowledge Base Update
Integrated with the company's content management system, when clauses or rules change, incremental index updates are triggered automatically, ensuring the knowledge base is always synchronized with the latest business rules.
Performance Data
| Metric | Before Optimization | After Optimization | Improvement |
|---|---|---|---|
| Knowledge Retrieval Time | 15 min | 10 sec | ↓99% |
| Answer Accuracy | 40% | 92% | ↑130% |
| New Hire Training Period | 3–6 months | 1–2 months | ↓67% |
| Knowledge Update Delay | 7 days | Real-time | ↓100% |
> Quantitative Summary: Knowledge retrieval time reduced by 99% to 10 seconds, answer accuracy improved by 130% to 92%, new hire training period shortened by 67% to 1–2 months, and knowledge update delay lowered from 7 days to real-time.
Technology Stack
Qwen2.5-72B, Milvus Vector Database, LangChain, FastAPI, Vue.js, Nginx Internal Network Deployment
FAQ
What recall rate can a RAG knowledge base typically achieve?
After optimization of chunking strategies and vector models, enterprise-grade RAG recall rates usually reach 85%–95%. In this case, answer accuracy improved from 40% to 92%, with the key being the quality of structured chunking and vector indexing of over 30,000 documents.
How does an insurance company's knowledge base meet the compliance requirement of keeping data within the domain?
All data processing and inference are completed within the internal network environment using a privately deployed Qwen2.5-72B model, with zero data transmission outside. The system meets the financial industry's compliance requirement of keeping data within the domain and has passed internal security audits.
Is the maintenance cost of a RAG knowledge base high?
Not high. The knowledge base is integrated with the company's content management system, automatically triggering incremental index updates when clauses or rules change, requiring no manual intervention. Daily maintenance involves only minor model tuning and troubleshooting, which can be covered by a single operations staff member.
