Building an Internal Knowledge MCP Gateway at Silicon2

Silicon2 built the S2 Domain Knowledge Gateway to integrate scattered internal knowledge (SP, source code, documents, etc.) into a form that AI can trust. This gateway uses MCP (Model Context Protocol) as the standard to connect models and tools.

During the development process, the AI-DLC methodology was applied, following these steps:

• On-prem/Embedded prototype: Established a foundation with a cost-efficient embedded structure for initial validation
• Migration to AWS managed services: After gaining internal standard approval, migrated state, audit, and deployment layers to AWS

For security, the principle of 'permission blocking before search' is strictly adhered to. Through an authorization layer utilizing AWS Cedar, the system is designed so that unauthorized users cannot see search results at all, and sensitive information is protected via field masking.

To improve search performance, hybrid search utilizing RRF (Reciprocal Rank Fusion) was introduced.

• Keyword search: Precise identifier search based on SQLite FTS5
• Vector search: Semantic search via the BGE-M3 embedding model