Developed and integrated ColBERT late interaction model support for document retrieval and ranking within the jeejeelee/vllm repository, focusing on enhancing document scoring accuracy and efficiency for large-scale corpora. Leveraged Python and machine learning techniques to implement per-token embeddings and MaxSim scoring, enabling improved ranking quality. The work included comprehensive API development, natural language processing, and thorough testing to ensure reliability and ease of adoption. Detailed documentation and example scripts were provided to guide users in utilizing the new model. This feature was delivered with an emphasis on code quality and production readiness, supporting robust validation and clear usage patterns.