Skip to content

honestAnt/wang_agents

Repository files navigation

English | 中文

Enterprise Agent Platform (EAP)

An enterprise-grade, multi-tenant AI Agent platform built on AgentScope. It is not just a chatbot — it is an Enterprise AI Operating System (AI OS) that unifies knowledge bases, model governance, agent runtime, skills, tools, memory, workflows, multi-tenancy, and AI observability into a single platform.

Target Users

  • Internal enterprise: AI platform teams, IT departments, digital transformation teams, data platform teams
  • SaaS scenarios: Intelligent customer service, enterprise knowledge assistant, digital employee, enterprise Copilot, lab AI assistant, workflow automation

Core Capabilities

Module Description
Enterprise RAG Document ingestion → parsing → chunking → embedding → hybrid search (BM25 + vector + metadata filter) with permission-aware retrieval
Intelligent Chat Multi-turn conversation, streaming output, multimodal input, agent collaboration, model switching
Agent Runtime AgentScope-powered execution loop (reason → act → observe), planner, router, executor
Skills System Prompt + Tool + Workflow capability encapsulation with dynamic loading, hot reload, and versioning
Tool Registry Unified enterprise tool hub supporting HTTP/MCP/SDK/Workflow tools with rate limiting, circuit breaking, and auditing
Model Gateway AgentScope-based multi-model governance (GPT, Claude, Gemini, Qwen, DeepSeek, vLLM, Ollama) with quota, budget, routing, and fallback
Memory System Short-term (Redis + sliding context) + Long-term (PostgreSQL + vector memory) — episodic, semantic, procedural
Multi-Tenancy tenant_id–based isolation at DB, index, and vector levels; RBAC + ABAC permission model
Trace & Audit Full-chain trace (User → Agent → Skill → RAG → Tool → LLM), OpenTelemetry + Langfuse, cost tracking, AI replay, prompt injection detection

Architecture

Frontend (Next.js)
  → API Gateway (Spring Cloud Gateway)
    → Java Microservices (Spring Boot — auth, user, agent, model, rag, tool, etc.)
      → Python Agent Runtime (AgentScope + FastAPI)
        → RAG Engine (LlamaIndex + Qdrant/Milvus)
        → Tool Mesh (MCP + internal APIs)
        → Model Gateway (AgentScope → LLM providers)
        → Trace (OpenTelemetry + Langfuse)

Data Layer: PostgreSQL / Redis / OpenSearch / Qdrant|Milvus / Kafka / MinIO

Tech Stack

Frontend

  • Next.js (TypeScript) — main site + admin console
  • Ant Design Pro — enterprise admin UI
  • Turborepo — monorepo management

Backend Services (Java)

  • Spring Boot microservices
  • Spring Cloud Gateway — API gateway
  • Keycloak — IAM (RBAC + ABAC)
  • PostgreSQL — primary OLTP database
  • Redis — session cache, short-term memory, rate limiting
  • Kafka — event bus
  • OpenSearch — full-text search, trace logs

AI Runtime (Python)

  • AgentScope — agent execution engine
  • AgentScope ModelWrapper — model gateway (unified API, fallback, routing, cost tracking)
  • LlamaIndex — RAG engine
  • Qdrant (dev/test) / Milvus (production) — vector database
  • Langfuse — LLM observability
  • OpenTelemetry — distributed tracing
  • Temporal / LangGraph — workflow orchestration (Phase 3)

Model Support

GPT-4.1, Claude, Gemini, Qwen, DeepSeek, vLLM, Ollama

Infrastructure

  • Kubernetes + Docker + Helm
  • MinIO / S3 — object storage
  • MCP (Model Context Protocol) — tool mesh standard

Project Structure (Planned)

enterprise-ai-platform/
├── docs/                  # PRD / architecture / API docs
├── deploy/                # k8s / helm / docker-compose
├── scripts/               # init scripts
├── frontend/              # Next.js monorepo
│   ├── apps/
│   │   ├── chat-ui/       # User chat workspace
│   │   ├── admin-console/ # Admin management
│   │   ├── agent-studio/  # Agent configuration studio
│   │   ├── model-center/  # Model governance
│   │   ├── rag-studio/    # RAG debugging & management
│   │   └── trace-console/ # Trace & audit console
│   └── packages/
│       ├── ui-components/ # Shared UI components
│       ├── api-client/    # Shared API client
│       ├── auth/          # Shared auth module
│       └── utils/         # Utility functions
├── backend-java/          # Spring Boot microservices
│   ├── common-lib/        # Shared library
│   ├── gateway-service/   # API Gateway
│   ├── auth-service/      # Keycloak IAM integration
│   ├── user-service/      # User/tenant/organization
│   ├── agent-service/     # Agent configuration
│   ├── model-service/     # Model governance
│   ├── rag-service/       # Knowledge base / RAG
│   ├── tool-service/      # Tool registry
│   ├── skill-service/     # Skills management
│   ├── memory-service/    # Long-term memory
│   ├── trace-service/     # Trace & audit
│   ├── billing-service/   # Cost billing
│   ├── prompt-service/    # Prompt engineering platform
│   ├── audit-service/     # Security audit
│   └── admin-service/     # Admin API aggregation
├── agent-python/          # AgentScope runtime (FastAPI)
│   └── app/
│       ├── core/          # Agent engine, planner, router, executor, memory, context
│       ├── agents/        # chat_agent, rag_agent, workflow_agent, multi_agent
│       ├── skills/        # data_analysis, customer_service, research
│       ├── tools/         # tool_client, mcp_client, internal_tools
│       ├── rag/           # retriever, reranker, embedding
│       ├── memory/        # short_term, long_term, vector_memory
│       ├── llm/           # model_wrapper, model_router
│       ├── trace/         # tracer, exporter
│       └── api/           # chat, agent, debug
├── sdk/                   # Internal SDKs
│   ├── java-sdk/
│   ├── python-sdk/
│   └── openapi/
├── infra/                 # Infrastructure configs
│   ├── mcp-servers/
│   ├── vector-db/
│   ├── search/
│   ├── message-queue/
│   └── observability/
└── shared/                # Shared protocols
    ├── proto/             # gRPC definitions
    ├── openapi/           # REST API schemas
    ├── event-schema/      # Kafka event definitions
    └── trace-model/       # Unified trace span schema

Development Roadmap

Phase 1 — MVP

Goal: Multi-tenant AI chat platform with basic RAG, tool calling, and multi-model switching

  • Monorepo setup + Docker Compose local infrastructure
  • Core Java microservices (auth, user, model, rag, tool, agent, gateway, trace)
  • Python Agent Runtime (AgentScope + FastAPI + LlamaIndex)
  • Frontend: Login, Chat UI, Dashboard, Agent/Model/KB management, Trace Console

Phase 2 — Advanced

Goal: Skills marketplace, long-term memory, multi-agent, prompt platform, full observability

  • Skill service + skill engine + skill marketplace
  • Long-term memory service (episodic/semantic/procedural)
  • Multi-Agent collaboration framework
  • Prompt Center (versioning, AB test, canary release)
  • Langfuse + OpenTelemetry full integration
  • Billing service

Phase 3 — Enterprise

Goal: AI OS — workflows, agent marketplace, intelligent routing, operations analytics, security governance

  • Temporal AI workflow engine
  • Agent marketplace
  • Intelligent model routing + auto fallback
  • AI operations analytics center
  • Security governance (prompt injection detection, data masking, privilege escalation detection)
  • SDK release + K8s production deployment + CI/CD

Getting Started (Prerequisites)

Local Development Dependencies

The .env file expects these services running locally:

Service Port Credentials
PostgreSQL 5432 opc_compliance / opc_123
MinIO 9000 minioadmin / minioadmin123
Redis 6379
Qdrant 6333
OpenSearch 9200

Quick Start (Coming Soon)

# 1. Start infrastructure
docker compose -f deploy/docker-compose.yml up -d

# 2. Initialize databases
bash scripts/init-local.sh

# 3. Start backend services
cd backend-java && ./mvnw spring-boot:run

# 4. Start agent runtime
cd agent-python && uvicorn app.main:app --reload

# 5. Start frontend
cd frontend && pnpm dev

Key Design Decisions

  • Trace standard: Every span MUST emit trace_id, span_id, type (llm|tool|rag|memory), latency, cost, tenant_id
  • Permission-aware RAG: Different users can only retrieve data they have access to, filtered by tenant_id, department, role, project
  • Data isolation: All SQL queries auto-inject tenant_id filter; vector and search indices are tenant-isolated
  • MCP-first tools: Model Context Protocol is the recommended standard for enterprise tool integration
  • Skills as first-class citizens: Prompt + Tool + Workflow encapsulation enables reusable AI capabilities across tenants

Current Status

Planning phase — architecture documents complete, code implementation not started.

See dev-map/ for design documents:

  • 企业智能体prd.docx — Full PRD
  • 企业智能体技术架构.docx — Technical architecture
  • 企业智能体页面设计.docx — UI/UX page design
  • 企业智能体工程目录实现.docx — Monorepo directory structure

Version

Current version: v0.1.0-alpha (Planning Phase)

See CHANGELOG for version history.

License

Copyright (c) 2026 fengbao.wang. All rights reserved.

  • Personal & Academic Use: Free — use, copy, and modify for personal projects and non-commercial academic research.
  • Commercial Use: Requires prior written authorization from the author. Contact wangfengbaowfb@gmail.com for commercial licensing.

See LICENSE for full terms.

Disclaimer

THIS SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR IMPLIED. THE AUTHOR SHALL NOT BE LIABLE FOR ANY DAMAGES ARISING FROM THE USE OF THIS SOFTWARE. USE AT YOUR OWN RISK.

About

基于agentscope构建企业级智能体平台(企业 AI Operating System(AI OS));包含 - 企业知识库 - 多模型治理 - Agent Runtime - Skills 管理 - Tool Registry - Memory - Workflow - 多租户权限 - AI 可观测性

Resources

License

Stars

Watchers

Forks

Releases

No releases published

Packages

 
 
 

Contributors