Copilot Assistant for Sage CRM

At the front of the architecture is a deterministic intent classifier that routes every request into one of three execution lanes: chit-chat, help, or data.

Chit-chat is handled entirely at the orchestration layer with predefined, personality-controlled responses. It never touches documents, models, or business systems, which keeps the system responsive and prevents unnecessary processing.

Help requests are routed into a local RAG pipeline. Documentation is pre-ingested, chunked, indexed, and ranked at runtime. Only the minimum relevant context is assembled and passed to the locally hosted model. A custom Modelfile enforces strict response constraints, ensuring answers remain grounded in the supplied context and never infer or invent citations.

Data requests are executed through a tightly controlled query layer against internal systems. Results are serialized, injected as bounded context, and explained by the model without allowing free-form interpretation, schema inference, or execution drift.

This separation enforces hard boundaries between conversational behaviour, documentation reasoning, and operational data access, reducing hallucinations, preventing data leakage, and keeping execution predictable — a requirement for regulated and security-sensitive environments.