Hybrid RRF Retrieval Beats Pure Vector Search for Kannada Literary RAG System

A developer building a retrieval-augmented generation (RAG) agent for 'Heli Hogu Kaarana', a 346-page scanned Kannada novel by Ravi Belagere, found that the initial pipeline using multilingual embeddings and ChromaDB produced confident but factually wrong answers. The core challenge stemmed from Kannada's agglutinative grammar, poor OCR quality from scanned pages, and sparse literary vocabulary that multilingual models handle poorly. The rebuilt system combined BM25 sparse retrieval with dense vector search, fused via Reciprocal Rank Fusion (RRF), and added a deterministic regex router to handle page-specific queries through direct metadata lookup. This hybrid architecture achieved 0.92 faithfulness and 0.89 context recall on a 50-query evaluation set, with a median end-to-end latency of 2.8 seconds on serverless infrastructure. The project concludes that retrieval architecture, not the language model itself, was the primary bottleneck in RAG performance for low-resource, morphologically complex languages.