Architecture — Loyalty Agent Service

System Overview

AI-powered assistant for managing a Loyalty Platform, built on the Model Context Protocol (MCP).

                 ┌──────────────────┐
                 │   React/Vite UI  │
                 │     :9333        │
                 └────────┬─────────┘
                          │ WebSocket / STOMP
                          ▼
                 ┌──────────────────┐
                 │  Loyalty Agent   │
                 │     :9332        │
                 │                  │
                 │ LLM Orchestrator │
                 │ MCP Client       │
                 │ Conversation     │
                 │ PostgreSQL       │
                 └────────┬─────────┘
                          │ MCP SSE (/sse, /message)
                          ▼
                 ┌──────────────────┐
                 │ Loyalty MCP      │
                 │ Server :9331     │
                 │                  │
                 │ 25+ MCP Tools    │
                 └────────┬─────────┘
                          │ REST + OAuth2 (Client Credentials)
                          ▼
                 ┌──────────────────┐
                 │ Loyalty Core API │
                 │     :8081        │
                 └──────────────────┘

          Keycloak ────── OAuth2 Token Provider
          PostgreSQL ──── Conversation Persistence
          Ollama ──────── Local LLM (Qwen3.5:4b)

Technology Stack

Technology	Version	Purpose
Java	25	Runtime
Spring Boot	4.0.0	Framework
Spring AI	2.0.0	LLM & MCP integration
Ollama	latest	Local LLM hosting
Qwen	3.5:4b	Primary LLM model
PostgreSQL	15+	Conversation persistence
Keycloak	latest	OAuth2/OIDC provider
React + Vite	19/6	Frontend UI

Module Architecture

loyalty-mcp-server (`:9331`)

MCP Server exposing Loyalty Platform capabilities as MCP tools via SSE transport.

Responsibilities:

Register 25+ MCP tools (campaign, rule, customer, transaction, etc.)
Validate tool inputs
Proxy requests to Loyalty Core API with OAuth2 client credentials
Return structured Result<T> with agent instructions for LLM formatting
Handle errors (sanitize stack traces, structured error responses)

Key Classes:

CampaignTools / CampaignRuleTools — Primary tool registration
CampaignService / CampaignRuleService — Business logic
RestClientConfig — OAuth2 REST client factory
Result<T> — Standard response wrapper with _agent_instruction

loyalty-agent (`:9332`)

AI Agent / BFF — Orchestrates LLM, MCP, conversations, and WebSocket.

Responsibilities:

Accept user messages via WebSocket/STOMP
Classify intent (regex-based, zero LLM latency)
Route to appropriate executor (SimpleAgent, Workflow, DirectChat)
Stream LLM responses back to frontend
Persist conversations to PostgreSQL

Key Architecture:

AgentController (@MessageMapping /chat)
        ↓
LoyaltyAgentService
        ↓
AgentOrchestrator
    ├── Active Workflow? → WorkflowExecutor (multi-turn creation)
    ├── IntentClassifier → QUERY → SimpleAgentExecutor (ReAct + tools)
    ├── IntentClassifier → CONVERSATION → SimpleAgentExecutor (direct chat)
    └── IntentClassifier → CREATE_* → WorkflowExecutor

Intent Classification (regex-based):

Intent	Trigger Examples
`QUERY`	"tìm chiến dịch", "danh sách rule"
`CONVERSATION`	"xin chào", "cảm ơn"
`CREATE_CAMPAIGN`	"tạo chiến dịch mới"
`CREATE_RULE`	"thêm thể lệ"

Tool Scope Enforcement:

Scope	Allowed Operations
`READ_ONLY`	search, get, find, count, check
`CREATE_CAMPAIGN`	READ_ONLY + createCampaign + generateCampaignId
`CREATE_RULE`	READ_ONLY + createRule + generateRuleId

Frontend (`:9333`)

React/Vite conversational UI with STOMP WebSocket.

Key Features:

STOMP over WebSocket for real-time messaging
Token streaming (character-by-character LLM output)
Tool execution status indicators
Conversation management (create, list, delete, rename)
Offline message queue with exponential backoff reconnect

Security Architecture

Frontend ──(WebSocket)──> Agent ──(MCP SSE)──> MCP Server ──(OAuth2)──> Core API
                                                     │
                                                     ▼
                                                  Keycloak

Principles:

LLM is never an authorization boundary — tool scope is enforced by AgentToolScope before tools reach the LLM
OAuth2 Client Credentials — MCP Server authenticates to Core API via Keycloak
Tool result sanitization — ToolInterceptor strips Java stack traces
Input validation — All tool parameters validated before downstream calls

Data Flow

Chat Request Flow

1. User types message in React UI
2. STOMP message → /app/chat → AgentController
3. AgentController → LoyaltyAgentService.chat()
4. AgentOrchestrator classifies intent
5. Executor starts LLM streaming with tools
6. LLM selects tool → ToolInterceptor wraps & calls MCP
7. MCP Server executes tool → calls Core API
8. Tool result → LLM reasons → generates response
9. Response streamed back via STOMP → /user/queue/chat-events
10. React UI renders tokens in real-time

Resilience Patterns

Pattern	Implementation
MCP auto-reconnect	`ToolInterceptor.attemptReconnect()`
WebSocket reconnect	`StompService` exponential backoff + jitter
Request timeout	`SimpleAgentExecutor` configurable via `agent.request-timeout-seconds`
Tool result truncation	`ToolInterceptor` caps results > 4000 chars
Error sanitization	`ToolInterceptor` strips stack traces
Graceful shutdown	`server.shutdown: graceful` on both modules
LLM search noise filtering	`SearchUtils.sanitizeSearch()`

6.6 KiB Raw Permalink Blame History