Need of retrieval augmented generation for conversational assistants for government social welfare scheme popularization

Governments allocate substantial resources to uplift under-privileged communities through various assistance schemes. Despite these efforts, the effectiveness of such initiatives often fails due to variety of reasons including cryptic nature of information available, lack of awareness to the needy people and or lack of transparency. In India, a vast and diverse developing democracy, these challenges are exacerbated by disparities in access to information. To address these issues, we explored the use of conversational AI technologies, including chatbots, which have seen successful application in many developed nations. Conversational AI agents often get challenged to a variety of factual inaccuracies and other types of hallucinations. Our study compares the accuracy of responses across different architectures: Our architectures integrate advancements in natural language processing, such as large language models (LLMs) and retrieval-augmented generation (RAG), along with automatic speech recognition (ASR), knowledge distillation, hybrid retrieval strategies, and text-to-speech systems. Overall this research experiments demonstrate that leveraging retrieval augmented generation using the same model embeddings substantially improves the quality of the output, its similarity to desired responses. This also reduces hallucinations. We also notice that larger models perform better than smaller parameter models.
This paper advocates for continued research in this field, suggesting that further advancements could substantially benefit those the schemes intend to serve.