TelcoGPT: A Hybrid Embedding Approach for Telecom-Specific Q&A and Code Retrieval

This paper presents TelcoGPT, a specialised question-answering (Q&A) and code retrieval system for telecommunications that combines retrieval-augmented generation (RAG) with domain-specific optimizations. TelcoGPT introduces three key enhancements: (1) a HybridEmbedding method-ology integrating text-embedding models with telecom-specific filtering mechanisms; (2) an advanced document processing pipeline with adaptive chunking and technical term density scoring; and (3) a dual-path query engine optimized for both question-answering and code retrieval tasks. Evaluation on the RedPajama-Data-1T arxiv subset demonstrates that hybrid embedding approach achieves mean reciprocal rank (MRR) of 0.89 and hit rate (HR) of 0.94 with optimal configuration (thresh-old=0.8, chunk size=12K, k=15), outperforming single-embedding approaches by 5-7%. The hybrid RAG implementation increases MRR by 8.5% (0.82 to 0.89) and HR by 10.6% (0.85 to 0.94). TelcoGPT achieves 95% accuracy in domain-specific Q&A tasks versus 87% for base models, while maintaining higher technical term density scores (0.90 vs 0.81). For code retrieval, our system demonstrates 93% execution success rate with comprehensive error handling, surpassing baseline approaches by 6-8%. Comparative analysis with GPT-3.5, GPT-4, and LLAMA-2/3 shows significant improvements in context relevance (0.92 vs 0.84), information accuracy (0.95 vs 0.89), faithfulness (0.84 to 0.92), and relevancy (0.83 to 0.93), demonstrating the effectiveness of our architecture for telecommunications applications.