Background 

A leading global manufacturer in North America sought to enhance customer experience insights through advanced sentiment analysis. With a diverse portfolio spanning building and consumer products, the company needed a scalable, secure, and cost-efficient solution to analyze reviews, extract contextual meaning, and deliver actionable intelligence. Their existing chatbot and analytics workflows lacked depth, leading to inconsistent outputs and delayed decision-making. 

Challenges 

The organization faced multiple roadblocks in deploying enterprise-grade AI: 

• Fragmented AI Operations: Data was being moved across multiple platforms, creating governance and security concerns. 

• Low Retrieval Precision: Traditional search methods delivered irrelevant results, limiting trust in insights. 

• High Costs and Latency: Legacy approaches generated slower response times and higher compute costs, reducing efficiency. 

• Quality Gaps: Chatbot responses often lacked correctness, completeness, and contextual accuracy, creating business risk. 

• Functional Limitations: No built-in session management, chat history, or visualization capabilities to support scale. 

Solution 

Narwal collaborated with the client to design and implement an Advanced Retrieval-Augmented Generation (RAG) Bot leveraging Snowflake Cortex and Streamlit: 

• End-to-End AI Operations Inside Snowflake 

• All embeddings, LLM generation, and orchestration executed within Snowflake for tighter governance and reduced data movement risk. 

• Hybrid Search and Re-Ranker 

• Combined metadata-driven hybrid search with a re-ranking mechanism to maximize contextual relevance. 

• LLM-as-a-Judge with Reflection 

• Integrated a feedback loop where an LLM (Llama3.1 & Sonnet based) validated outputs for correctness and completeness, reducing hallucinations and ensuring trustworthiness. 

• Reasoning Model Integration 

• Enhanced the generation pipeline with reasoning capabilities, boosting depth and logical consistency of responses. 

• Optimized Embeddings and Execution 

• Adopted the snowflake-arctic-embed-l-v2.0 embedding model 

• Achieving ~70% retrieval precision. 

• Response quality with ~77% consistency and ~68% completeness across multiple test runs. 

• Adopted the snowflake cortex search service 

• Achieving ~83% improved retrieval precision. 

• Increased response quality with ~93% consistency and ~83% completeness across multiple test runs. 

Outcomes 

The modernization effort yielded measurable results: 

• Faster Processing: Achieved ~29% lower latency compared to legacy GPT-based paths (≈21.9s vs 30.6s). 

• Cost Efficiency: Significantly reduced generation and ingestion costs, enabling sustainable scale. 

• Enhanced Security: All AI operations consolidated within Snowflake, minimizing external data transfers and exposure. 

• Improved Accuracy: Achieved ~83% retrieval precision with high response consistency, strengthening business trust. 

• Proof-of-Concept Success: Delivered a production-ready RAG chatbot prototype with Streamlit UI, validated through MLQA checks (Precision@K, consistency tests, and multi-answer validation). 

Conclusion 

Narwal’s advanced RAG bot implementation transformed the client’s sentiment analysis capabilities, delivering faster, cheaper, and more accurate insights—all while ensuring enterprise-grade security and governance. By combining hybrid search, reasoning models, and LLM-as-a-Judge pipelines, the solution set a strong foundation for scaling AI-powered customer engagement and analytics across the enterprise. 

Partner with Narwal today to modernize your AI operations and unlock business-ready intelligence with security and scale.