Swisscom: RAG Assistant for Customer Support

The challenge

Swisscom's assistant 'Sam' had three critical problems: it answered in the wrong language, provided irrelevant or broken URLs, and gave responses too short, pushing customers to request a human agent.

Our solution

We built a RAG architecture with LangChain featuring two-phase retrieval: similarity search (K=30) followed by BM25 reranking (K=3). Embeddings with OpenAI text-embedding-3-large, ChromaDB vector store. Generation with GPT-4o with conversation history summarized at each turn. Native multilingual support (DE/FR/IT/EN) without additional filters, thanks to chunk-level reranking.

Key results

• Average score of 0.78 on scale vs Sam across 85 questions in 4 languages
• An impartial LLM judge prefers our assistant 8 out of 10 times
• Language detection problem completely solved
• Relevant and verified URLs in responses

Related case studies