A Kannada centric retrieval augmented generation system for Karnataka heritage knowledge access

Abstract Large Language Models produce fluent text, but in domain-specific and low-resource language settings they are prone to hallucination, weak factual grounding and limited multilingual coverage. These problems are acute for Kannada heritage information, where historical sources are fragmented, sparsely digitised and often available only as scanned documents. The presented work is on Retrieval Augmented Generation system for Kannada question answering over Karnataka cultural heritage documents. The system combines hybrid document ingestion using PyMuPDF and Tesseract OCR with per-page provenance, Kannada-aware Unicode normalisation, recursive chunking on a Kannada-aware separator hierarchy, a general-purpose multilingual sentence encoder, and exact cosine retrieval over a FAISS index. A deterministic passage selection rule combining an absolute and a relative similarity floor sets the number of retrieved passages from the score distribution, and the system abstains when no passage clears the absolute floor. Retrieved passages are supplied to locally deployed LLMs through FastAPI and Ollama with an instruction to answer only from them, enabling fully offline operation, each response lists the source page of every passage supplied to the model. The system was evaluated on 25 Kannada questions written by the research team (8 factual, 9 interpretive and 8 analytical) over a single 62-page heritage document. On this set, gemma3:4b obtained the highest scores on all four response-quality metrics and the lowest mean latency (3.97 s), with BERT-F1 of 0.850, 0.863 and 0.851 across the three tiers (0.855 aggregate). Across all models, BERT-F1 remained high and nearly flat across tiers, while ROUGE and alignment scores fell sharply on interpretive and analytical questions. We analyse this divergence and argue that, in a morphologically rich low-resource language, neither family of metrics is self-validating without retrieval-level and human evaluation. All reported tables are derived from a single per-question record log with automated consistency checks. The system offers a starting point for offline Kannada heritage question answering whose retrieval and grounding remain to be validated.