Enterprise service platforms connect various knowledge artifacts and office applications in organizations to enable automation of routine decision-making. During this automation, service requests are expressed as domain-independent knowledge queries to capture gaps in knowledge related to governance, operations, risk management, customer service, and other enterprise aspects, and stored in a knowledge repository. Retrieval-augmented generation driven by small-scale pre-trained transformers offers an ideal means to automate responses to such queries because information retrieval and text-to-text generation can be achieved using state-of-the-art—if not better—large language models without incurring the high inference costs associated with their larger counterparts. A system architecture providing this functionality is presented, together with an exploration of the elements of the knowledge-retrieval phase. Empirical evaluation of the effectiveness of the retrieval step shows that it satisfies the requirements of a diverse set of queries.

Deployments of enterprise service platforms within organizations have shown that a significant proportion of service requests relate to knowledge gaps in domains such as governance, operations, risk management, customer service, and so on. Efforts to support automation of these decision-making tasks attempt to address such requests by posing knowledge-retrieval queries for the pertinent answers. Cross-domain databases, policy repositories, internal and external knowledge bases, and other such information collections serve as knowledge sources. To support these requests, retrieval-augmented generation leverages a combination of information retrieval and large language models.