"Knowledge Retrieval Based on Generative AI"

This paper introduces a RAG-based system combining BGE-M3 embedding and BGE-reranker to enhance LLM accuracy while maintaining data privacy through local infrastructure deployment.

-----

https://arxiv.org/abs/2501.04635

🔧 Solution in this Paper:

→ The system uses BGE-M3 for dense vector retrieval across 100+ languages with 8,192 token capacity.

→ BGE-reranker prioritizes query-relevant results through cross-encoding.

→ Chinese Wikipedia (1.38M entries) and Lawbank financial regulations serve as knowledge sources.

→ FAISS implements efficient vector indexing using Flat Index, IVF, and HNSW methods.

→ Local infrastructure deployment ensures data privacy and reduces commercial service dependence.

-----

💡 Key Insights:

→ Domain-specific knowledge sources outperform general sources for specialized queries

→ Re-ranking significantly improves result relevance compared to semantic similarity alone

→ Local deployment enhances privacy while maintaining performance

→ Human assistance with RAG dramatically improves accuracy in specialized domains

-----

📊 Results:

→ Taiwan-LLM-8x7B-DPO accuracy: 57.28% to 88.35%

→ ChatGPT 3.5 accuracy: 74.76% to 88.35%

→ Human performance improved from 28.75 to 85.25 average score

→ Financial banking questions showed 13-16% accuracy improvement