AI Layer

The AI layer provides reasoning and knowledge retrieval capabilities within the Demiton platform.

Its role is to assist users in understanding operational data and documents.

However, the AI layer does not execute actions directly.

All system mutations occur through Blueprint workflows in the execution runtime.

This separation ensures that AI reasoning never bypasses deterministic infrastructure.

Responsibilities

The AI layer performs several functions.

Natural Language Interaction

Users can interact with operational data using natural language queries.

Examples include:

• "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.

This process works as follows:

1. user submits a question
2. the system performs vector search
3. relevant document fragments are retrieved
4. fragments are inserted into the model prompt
5. the model generates a response

This ensures responses remain grounded in enterprise data.

Document Intelligence

The AI layer can analyze operational documents including:

• site diaries
• project reports
• safety logs
• uploaded documents

These documents are indexed and stored in vector memory for retrieval.

Architecture Components

The AI system consists of several components.

MCP Service

The Model Context Processing service orchestrates AI interactions.

Responsibilities include:

• processing chat turns
• assembling context
• invoking the model provider
• storing conversation history

LLM Service

The LLM service provides controlled access to language models.

Responsibilities include:

• model invocation
• prompt assembly
• token configuration
• provider integration

Vector Memory

Vector storage enables semantic retrieval.

Indexed documents contain:

• normalized text content
• embedding vectors
• metadata
• access control fields

Security Model

AI retrieval follows strict security rules.

Every retrieval request includes an identity filter.

Documents are indexed with:

allowed_entra_users
allowed_entra_groups

The system only retrieves documents the user is authorized to access.

This ensures the model never receives unauthorized data.

Execution Boundary

AI output is advisory only.

The model cannot:

• mutate ERP systems
• call adapters
• execute financial operations

If a user requests an operational action, the AI layer may suggest a Blueprint workflow.

Execution only occurs after explicit user approval.

Design Principle

The AI layer enhances clarity and productivity.

Infrastructure remains responsible for execution.

This separation ensures reliable system behavior in enterprise environments.

---