RAG 2.0: Gen AI in Advanced Supply Chain Analytics

The Department of Defense's supply chain management ecosystem faces unprecedented challenges in maintaining military readiness and operational efficiency due to increasing global complexity and rapid technological advancement. Current approaches rely heavily on manual analysis by large teams, resulting in decision-making delays and reduced operational agility. This proposal presents an innovative solution that will leverage advanced Retrieval-Augmented Generation (RAG) 2.0 technology to transform how the Defense Industrial Base processes and acts upon supply chain information. The proposed system will be able to convert queries and varied, complex data sources into highly actionable and informative responses. Our system will implement state-of-the-art RAG 2.0 technologies and models to introduce useful features like matching claims to source documents, as well as guarantee industry leading levels of accuracy. By utilizing a unique mixture-of-retrievers framework and by placing an emphasis on the unified overall system rather than specific plug and play models, we will be optimally suited to satisfy the Air Force’s high standards. The foundation of the proposed system will be a sophisticated document ingestion and extraction system specifically tuned to excel within the supply chain domain. Our solution will allow documents of all variations to be ingested en masse into our system. These documents will enter into an extraction queue and be converted into chunks to be stored within a vector database. User queries will be similarly encoded and used to run hybrid search algorithms to acquire the relevant chunks. The retrieved chunks will pass through several retriever models to ensure maximum relevancy and accuracy. The chunks will first flow through our industry leading reranker model followed by an LLM-based filtering stage. This condensed set of thoroughly vetted chunks will then be passed with the initial query into our specifically tuned LLM and used to generate a response. The response will undergo claim delineation and contextualization followed by an attribution process to help shape the answer and connect claims to the relevant chunks. A final groundedness check will perform the last round of validation to ensure that each claim is accurately backed up by the support chunk attached to it. This state-of-the-art system of retrievers and models will guarantee unprecedented accuracy and maximize the utility of the responses. Our world class research and development teams, led by one of the original authors of the foundational RAG paper, Douwe Kiela, have already made huge strides in the initial underlying research. Contextual AI is fully prepared to execute on the proposed solution and develop the world class RAG 2.0 solution the Air Force is searching for. Our expertise in generative AI models, combined with our emphasis on enterprise needs, makes us optimally suited to help modernize supply chain analytics for the Air Force.