Dispatching the Parsed RAG Question: Chunk Strategy, Model Tier, Activations, Audit

Dispatching in a Retrieval-Augmented Generation (RAG) system involves making decisions after parsing a user's question by using the document's profile. This process determines the optimal chunking strategy, the amount of context to retrieve, and which language model from a tiered system is most appropriate for the query. The system uses satellite tables that define rules based on the question's shape, type, and conceptual category to make these choices deterministically. This allows for more efficient and context-aware responses, such as using a small model for a simple fact extraction or a powerful one for a complex summary. Ultimately, this approach moves beyond simple keyword retrieval to a more intelligent system that optimizes for both cost and performance.