Architecture for a Future Model
The interfaces prepared for a future language-model layer — the educational, research, and expert answer modes, the RAG-ready interfaces, prompt orchestration, conversation memory, and LLM integration. Defined and modelled; a model, when added, would phrase the grounded facts and never add to them.
- Answer ModesArchitectureArchitecture-ready audience modes
An architecture-ready interface for tuning the depth of an answer to its audience — an educational mode for newcomers, a research mode for depth, an expert mode for concision. The modes are defined; a language-model layer would realise them over the same grounded facts.
- Conversation MemoryArchitectureArchitecture-ready session memory
An architecture-ready interface for remembering the thread of a conversation — the entities discussed, the questions asked — so follow-ups stay in context. The interface is prepared; memory stays on the device, private-first, and holds only references into the graph.
- LLM IntegrationArchitectureArchitecture-ready model layer
The seam where a language model would join — phrasing the grounded facts the retrieval layer supplies, always constrained to what the graph can back. The interface is defined; the model is future work, and when it comes it will explain, never invent.
- Prompt OrchestrationArchitectureArchitecture-ready orchestration
An architecture-ready layer for orchestrating a language model over the grounded retrieval — deciding what to fetch, in what order, and how to assemble it. Defined as an interface; no model is wired in yet.
- RAG-Ready InterfacesArchitectureRetrieval-augmented-generation interface
The grounded retrieval — search, neighbourhoods, evidence chains — is exposed as a retrieval-augmented-generation interface, so a future language model can ground every response in the real graph. The retrieval is real today; the generation is future work.