The problem
Employees spend significant time searching through fragmented PDFs, Notion pages, and Drive directories with poor keyword-based search relevance.
The approach
Implemented a dense-sparse hybrid retrieval pipeline (BGE-M3 + BM25) combined with Cohere Rerank. Built a self-correcting RAG architecture that performs citation verification and hallucination detection before answering.
Results
Achieved 94% answer accuracy on user-validated benchmarks, reducing time-to-information for operations staff by 70%.