Why Pure Vector Search Fails on Kannada Literature — And How Hybrid RRF Fixed It

Pure Vector Search struggles on Kannada literature due to retrieval issues. The RAG (Retrieval-Augmented Generation) application faced problems stemming from:

1. Scanned physical book: No digital source; OCR accuracy is a challenge.

2. Kannada's agglutinative nature: Names like "Himavant" can appear in various inflected forms, making it difficult for multilingual embeddings to accurately represent semantic meaning.

3. Sparse and specific literary text: Rare colloquialisms, proper nouns, and page-level references are easily missed by generic embeddings trained on high-resource data.

The initial RAG pipeline used chunking, embedding with a multilingual model, storing in ChromaDB, retrieving top-5 results, and prompting Gemini. However, this approach failed in several ways:

1. Proper nouns were frequently missed.

2. Rare or inflected forms returned thematically similar but factually incorrect chunks.

3. Page queries became "vibe" queries instead of precise answers.

These issues led to wrong answers with full confidence, and the system had no mechanism to detect its mistakes.

The architecture improved by introducing a deterministic regex router that bypasses semantic search for exact-page queries, achieving 100% precision and zero hallucination. The final system achieved 0.92 RAGAS faithfulness and 0.89 context recall on a 50-query golden set, with a 2.8s P50 end-to-end processing time on serverless infrastructure.