Secure RAG: Bridging LLMs and Local Deployment for Corporate Intelligence Search

Below is the English abstract for your submission, refined to reflect your technical depth and strategic engineering choices:

Abstract The real-world deployment of RAG (Retrieval-Augmented Generation) often faces challenges from corporate security policies and hardware constraints. This session uses the Mohuhu platform as a case study to explore how open-source LLMs and system integration techniques solve complex semantic relationships and search fault-tolerance issues for Taiwan’s listed company data.

We will deconstruct how AI capabilities are injected into different system layers:

LLM-Driven Index Enrichment: Using Google Gemini as a knowledge agent to automatically generate unstructured aliases (e.g., "護國神山," "2330") during the indexing stage, fundamentally addressing semantic gaps in vector search for domain-specific data.

Four-Strategy Hybrid RAG: The system’s core is a weighted scoring engine that aggregates confidence scores across four dimensions:

Exact Match: Ensuring absolute accuracy for stock codes and tax IDs.

Chinese Fuzzy Search: Handling character misplacements and redundant terms.

Phonetic/Bopomofo Fault-Tolerance: Solving search pain points caused by incorrect input methods (e.g., typing vm6vm42j/).

Semantic Vector Search: Integrating local Ollama (bge-m3) and pgvector to process deep semantic connections.

Deployment Trade-offs for High-Security Environments: Proposing an "AI Pre-processing + Lightweight Local Retrieval" architecture for offline scenarios. Using PyInstaller for encapsulation, allowing enterprises to enjoy LLM-enhanced experiences in completely offline environments.

Automated LLM Workflows: Utilizing OpenClaw and Telegram Bots to monitor the ETL pipeline—from crawling to embedding—and implementing automated retry mechanisms for failures.

Key Takeaways Hybrid Search Implementation: Detailed design of a four-way parallel weighted architecture and weight-tuning strategies.

RAG Practical Strategies: Improving retrieval recall in specific domains using LLM-based data pre-processing.

Engineering Design Patterns: Practical experience in balancing data security, hardware limitations, and AI performance during deployment.