Enterprise AI Knowledge Platform

A secure internal AI layer for a large Swiss enterprise: model-agnostic agents, SharePoint and Elasticsearch ingestion, vector + graph retrieval, MCP-based permissions, local GPU inference, cloud fallback, and anonymization.

Enterprise AI architecture

Large Swiss Enterprise

Permission-aware RAG

The Challenge

A large enterprise in Switzerland needed an AI system that could answer business-specific questions across internal documentation, historical tickets, SharePoint files, search indexes, and department-specific operational knowledge.

The challenge was enterprise-grade retrieval, not a simple chatbot: the platform had to respect user permissions, combine local and cloud models, protect sensitive data, and keep working even when one knowledge layer returned incomplete results.

Fragmented

Enterprise knowledge

Strict

Permission boundaries

Hybrid

Local + cloud AI

Our Solution

We supported a modular AI knowledge platform that routes requests between specialized agents, internal tools, local GPU-hosted models, and selected cloud models depending on the task and data sensitivity.

The architecture combines semantic vector search with graph-based relationship retrieval, connected through MCP servers that enforce permissions before data reaches the model. A custom anonymization layer masks sensitive information when cloud processing is useful.

Technology Stack

AI, Retrieval & Data

Local LLMs with GPU inference
Cloud LLM routing for selected workflows
LanceDB for vector search
Neo4j / graph database retrieval
PostgreSQL for traces, memory, and audit data

Enterprise Integration & Security

MCP servers for tool orchestration
SharePoint ingestion connector
Elasticsearch MCP integration
Dockerized deployment on controlled infrastructure
Prompt-injection guardrails and anonymization

Vector + Graph Retrieval

Semantic search and relationship-based reasoning were kept separate, giving the platform fallback resilience and better answers for complex internal knowledge.

Permission-Aware MCP Layer

MCP servers enforce user-specific access rules before agents query internal systems, keeping finance, HR, operations, and restricted data properly scoped.

SharePoint + Search Ingestion

Connectors track document updates, ingest useful ticket history, and integrate Elasticsearch so agents can retrieve knowledge from existing enterprise systems.

Anonymization + Guardrails

Sensitive identifiers can be masked before cloud model calls, while moderation and prompt-injection checks reduce risk across enterprise AI workflows.

The Results

The platform created a secure foundation for internal AI assistants that can retrieve knowledge across enterprise systems, respect access controls, and choose the right model or tool for each task.

Secure

Enterprise Retrieval

Permission-aware by design

Hybrid

Model Strategy

Local GPU + cloud routing

Resilient

Knowledge Layer

Vector and graph fallback

Architecture note

The most important design choice was to keep retrieval, permissions, model routing, and anonymization as separate layers. That makes the system easier to audit, safer to extend, and more resilient when one source returns incomplete or outdated information.

The Process

Knowledge & Access Mapping

Mapped internal sources, user roles, sensitive data boundaries, document update patterns, and the types of questions employees needed to answer.

Data Source MappingPermission DesignUse Case Scoping

Retrieval & Tooling Architecture

Implemented vector and graph retrieval, MCP tool boundaries, SharePoint ingestion, Elasticsearch access, trace analysis, and model-routing logic.

MCP ToolsVector RAGKnowledge Graph

Secure Deployment & Iteration

Containerized the platform, tuned retrieval quality through execution traces, added anonymization and guardrails, and prepared the system for enterprise-controlled infrastructure.

Docker DeploymentGuardrailsQuality Tuning