Agentic Tool-Augmented Translation for Low-Resource Sanskrit--English under Resource Constraints

Sanskrit remains challenging for machine translation not only because of its low-resource setting, but also because of its rich morphology, culturally specific vocabulary, and dense historical and religious references. General-purpose large language models often struggle with these characteristics, producing mistranslations of culturally embedded meanings, ritual concepts, and historical genealogies. We propose a training-free, tool-augmented framework that grounds LLM-based Sanskrit–English translation in external linguistic and cultural resources. The framework orchestrates Monier–Williams dictionary lookup, morpho-syntactic analysis, retrieval of relevant translation examples, and glossary constraints at inference time, while keeping the underlying language model frozen. We evaluate the approach on five Sanskrit–English datasets spanning classical epic and poetic text, religious prose, political speech, technical instruction, and pedagogical material. Across these domains, the proposed agentic pipeline consistently improves BLEU and chrF over the frozen LLM baseline, with qualitative analyses showing better lexical disambiguation, terminology consistency, and preservation of culturally specific entities and expressions. These results suggest that explicit linguistic grounding can compensate for some limitations of parametric knowledge in low-resource translation without requiring model fine-tuning.