Vector Search Service (MCP Provider)

Backend API for data registration and semantic search, acting as an MCP (Model Context Protocol) context provider.

Tech Stack

• Node.js + TypeScript (Strict mode)

• NestJS (HTTP Framework and Dependency Injection)

• PostgreSQL + pgvector (Vector Database)

• Prisma (ORM)

• OpenAI / Gemini / Claude (Voyage AI) (Interchangeable Embedding Adapters)

• Helmet + Throttler (Security headers and Rate Limiting)

Related MCP server: ragkit

Architecture

The project follows Clean Architecture and Hexagonal Architecture:

• Domain: Framework-agnostic Entities and Interfaces (Ports).

• Application: Use Cases containing business rules.

• Infrastructure: Concrete implementations (Prisma, AI Adapters).

• Interface: HTTP Controllers (REST).

Configuration

1. Installation and Environment

1. Install dependencies:

   npm install

2. Configure environment variables: Copy .env.example to .env and fill in the keys.

   cp .env.example .env

   Supported Adapters (EMBEDDING_PROVIDER):

   • openai: Requires OPENAI_API_KEY (Model: text-embedding-3-small) (Recommended)

   • gemini: Requires GOOGLE_GENAI_API_KEY (Model: gemini-embedding-001)

   • claude: Requires ANTHROPIC_API_KEY and VOYAGE_API_KEY (Model: voyage-large-2).

Security Configuration:

• MAX_CONTENT_ITEMS: Maximum number of items allowed in the data array (Default: 100).

2. Database with Docker

You can start the PostgreSQL database with pgvector extension using Docker Compose:

docker compose up -d

This will start the database on port 5432. Default credentials are in docker-compose.yml.

3. Database Migration

Run the Prisma migrations to create the database schema:

npx prisma migrate dev

4. Execution

To start the application in development mode:

npm run start:dev

The application will be available at http://localhost:3000.

5. Swagger Documentation

Interactive API documentation (Swagger UI) is available at:

http://localhost:3000/api

API Endpoints

1. Register Data (Upsert)

POST /data/register

Registers data for a project and content ID. Replaces existing data (idempotent).

Payload:

{
  "projectId": "project-alpha",
  "contentId": "doc-123",
  "data": [
    "MCP allows connecting AIs to external data.",
    "Vector search uses embeddings for similarity."
  ]
}

Response (201):

{
  "message": "Data registered successfully"
}

2. Search Data

POST /data/search

Semantic search on registered data.

Payload:

{
  "search": "how does vector search work?",
  "projectId": "project-alpha",
  "limit": 3
}

Response (200):

{
  "results": [
    {
      "projectId": "project-alpha",
      "contentId": "doc-123",
      "data": [
        "Vector search uses embeddings for similarity."
      ]
    }
  ]
}

3. List Data (Grouped)

GET /data

Lists stored data, grouped by project and content, returning item counts. Supports pagination.

Query Params:

• projectId (optional): Filter by project.

• contentId (optional): Filter by content ID.

• page (default: 1): Page number.

• limit (default: 10): Items per page.

Response (200):

{
  "results": [
    {
      "projectId": "project-alpha",
      "contents": [
        {
          "contentId": "doc-123",
          "items": 2
        }
      ]
    }
  ],
  "page": 1,
  "limit": 10
}

4. Remove Data

DELETE /data

Removes data filtering by project or content ID. At least one filter is required.

Query Params:

• projectId: Project ID.

• contentId: Content ID.

Response (200):

{
  "message": "Data removed successfully"
}

Security Features

• Input Validation: Strict DTO validation with class-validator (whitelist enabled).

• Rate Limiting: Global rate limiting (100 requests/minute) using @nestjs/throttler.

• HTTP Headers: Secure HTTP headers via helmet.

• CORS: Enabled with default settings.

• SQL Injection Protection: Uses Prisma's parameterized queries and raw SQL template literals for vector operations.

• Payload Limits: Configurable limit for input array size (MAX_CONTENT_ITEMS).

Tests

E2E Tests

npm run test:e2e

Unit Tests

npm run test

Folder Structure

src/
├── application/       # Business Rules (Use Cases)
├── domain/            # Entities and Interfaces
├── infrastructure/    # Implementations (DB, Adapters)
├── interface/         # Controllers and DTOs
└── main.ts            # Entrypoint