Foundgine

AI Integration

Foundgine is AI-native at the application boundary, but it is not an AI model framework.

Position

LLM / Agent framework
        │
        ▼
      Intent
        │
        ▼
    Foundgine
        │
        ├── semantic resolution
        ├── policy
        ├── planning
        ├── execution
        ├── verification
        └── evidence

The external AI system owns reasoning and conversation.

Foundgine owns application semantics and safe execution.

Why this boundary matters

A generic AI system does not automatically know:

which entities exist
which relationships are legal
which operations are business actions
which actions mutate state
who may perform them
how to verify the result

Those facts belong to the application.

AI integration principles

1. Constrained vocabulary

The model should select from a known semantic vocabulary.

2. No arbitrary method invocation

Foundgine must never translate an LLM-generated method name into arbitrary CLR invocation.

3. Plan before mutation

Mutations should become explicit plans before execution.

4. Policy before execution

Authorization is part of planning.

5. Evidence after execution

The agent should receive structured evidence rather than trusting an unverified textual claim.

MCP

MCP is a planned outer adapter:

Agent
 ↓
MCP
 ↓
Foundgine semantic API

Foundgine does not need to implement the MCP protocol inside its core.

Retrieval

Semantic retrieval is a planned execution target, not the definition of Foundgine.

A future system can combine:

Structured data
+
Semantic retrieval
+
Domain actions
+
External data

under the same planning/evidence model.

AI accuracy rules

Never claim that Foundgine currently has:

a complete agent framework
natural-language planning in production
complete MCP support
complete RAG support
autonomous mutation safety
enterprise-grade authorization
production-ready evidence/audit

Those are roadmap capabilities until demonstrated by code and tests.

See:

This site is open source. Improve this page.