Service composition, the process of combining existing services to create new functionalities, represents a crucial aspect of Service-oriented Computing (SOC). Large Language Models (LLMs) have been shown effective for the task of automatic service composition, however the limited context size poses a significant technical limitation as it limits the number of service descriptions that can be included within the prompt. Additionally, not all existing work captures and bundles service descriptions used for service composition in a standardized format such as OpenAPI. When evaluating the performance of a service composition approach the underlying dataset used often utilizes query-gold solution path tuples as the determining factor for a successful composition. The use of a single per query solution path can falsy deem correct compositions as unsuccessful if multiple valid solution paths exist. This work investigates the application of Retrieval Augmented Generation (RAG) frameworks to address limited context sizes, focusing on dataset construction and retrieval optimization. We implement a LlamaIndex pipeline allowing developers to ingest OpenAPI specifications into a vector store and provide an OpenAPI parser to reduce the token counts of OpenAPI documents. Further, we provide an extended retriever that processes queries using query transformation modules and applies reranking to embedding retrieval. We analyze an exemplary extended retriever pipeline measuring the Normalized Discounted Cumulative Gain (NDCG), recall, precision and F1 score. Lastly, we implement an entity-based evaluation process aimed to gather service endpoint responses necessary to fulfill a query while reducing the number of necessary calls.