Language models as a service (LMaaS) enable users to accomplish tasks without requiring specialized knowledge, simply by paying a service provider. However, numerous providers offer massive large language model (LLM) services with variations in latency, performance, and pricing. Consequently, constructing the cost-saving LLM services invocation strategy with low-latency and high-performance responses that meet specific task demands becomes a pressing challenge. This paper provides a comprehensive overview of the LLM services invocation methods. Technically, we give a formal definition of the problem of constructing effective invocation strategy in LMaaS and present the LLM services invocation framework. The framework classifies existing methods into four different components, including input abstract, semantic cache, solution design, and output enhancement, which can be freely combined with each other. Finally, we emphasize the open challenges that have not yet been well addressed in this task and shed light on future research.
View on arXiv