AI Layer
The AI layer provides reasoning and knowledge retrieval capabilities within the Demiton platform. It assists users in understanding operational data and documents. It does not execute actions directly — all system mutations occur through deterministic workflows in the execution runtime.
This separation is intentional. AI reasoning should inform decisions; infrastructure should execute them.
Responsibilities
Natural language interaction — users can query operational data using natural language:
- “What happened on Project Alpha yesterday?”
- “Summarize this week’s site activity.”
- “Show labour hours for Project Delta.”
Knowledge retrieval — the platform uses retrieval-augmented generation (RAG):
- User submits a question
- Vector search retrieves semantically relevant document fragments
- Fragments are assembled into the model context
- The model generates a grounded response
Document intelligence — the AI layer can analyze operational documents: site diaries, project reports, safety logs, uploaded files. These are indexed and stored in vector memory for retrieval.
Architecture components
MCP service — the Model Context Protocol service orchestrates AI interactions. It processes chat turns, assembles context from multiple providers, invokes the model, and stores conversation history.
Context providers — specialized providers fetch relevant data for each conversation: project state, worker records, workflow history, connected system data.
LLM service — provides controlled access to language models. Manages model invocation, prompt assembly, token configuration, and provider selection (Azure AI Foundry or Claude Enterprise).
Vector memory — Azure AI Search stores document embeddings. Every retrieval query applies identity-based filters — documents only return to users who are authorized to access them.
Security model
Every retrieval request includes an identity filter. Documents are indexed with allowed_entra_users and allowed_entra_groups fields. The vector search query always scopes results to the requesting user’s identity.
The model never receives documents the requesting user is not authorized to see.
Execution boundary
The AI layer is advisory only. The model cannot:
- call adapters
- mutate ERP systems
- execute financial operations
- trigger workflows autonomously
If a user requests an action, the AI layer may suggest a workflow. Execution only occurs after explicit user approval in the platform UI.
Providers
| Tier | Provider |
|---|---|
| Default | Azure AI Foundry |
| Enterprise | Claude Enterprise |
Provider selection is configured per-tenant based on tier and connector configuration.