China Unicom's open-source enterprise-grade multi-tenant AI agent platform, integrating Model Hub, Visual Workflow, High-precision RAG, MCP toolchains, and Xinchuang compatibility.
Wanwu (元景万悟) is China Unicom's enterprise-grade, multi-tenant AI agent development platform under Apache 2.0 license, deployed across finance, industry, and government sectors.
Core Capabilities#
Model Management (Model Hub): Unified access and lifecycle management for hundreds of commercial/open-source LLMs (GPT, Claude, Llama, etc.), with deep OpenAI API compatibility and support for vLLM, TGI inference backends and private deployment.
Visual Workflow (Workflow Studio): Low-code drag-and-drop canvas for complex business processes with conditional branching, API, LLM, knowledge base, code, and MCP nodes, supporting end-to-end debugging and performance analysis.
Agent Framework: Function Calling-based agent construction with tool extension, private knowledge base association, multi-turn dialogue, and online debugging support.
High-precision RAG: Full pipeline from knowledge base creation → document parsing → vectorization → retrieval → fine ranking. Supports pdf/docx/txt/xlsx/csv/pptx and web resources; integrates multimodal retrieval, cascade and adaptive segmentation, GraphRAG (UniAI-GraphRAG). On MultiHop-RAG dataset, Wanwu GraphRAG outperforms Dify by ~17.2%, LightRAG by ~3.5%.
Tools & Connectivity#
MCP (Model Context Protocol): Standardized interface for external tool integration with 100+ built-in industry MCP interfaces; MCP Hub supports importing MCPs from various providers.
Web Search: Real-time information acquisition, multi-source data integration, intelligent search strategies (keyword/semantic retrieval).
External Knowledge Base Compatibility: Supports API-based import of Dify-created knowledge bases.
Platform & Services#
Backend as a Service (BaaS): RESTful API with fine-grained permission control.
Security & Publishing: Sensitive word lists for output safety control; public/private publishing options for text Q&A/workflows/agents as API or Web URL.
OpenClaw Sandbox: Optional deployment of OpenClaw Robots in isolated Docker containers.
Architecture & Xinchuang Compatibility#
- Enterprise multi-tenant architecture for cost control, data isolation, elastic scaling
- Modular design for flexible extension and secondary development
- Xinchuang compatibility: Kunpeng CPU, openEuler/CULinux/Kylin, TiDB/OceanBase
- Certified with "Xinchuang AI Software-Hardware System Detection Certificate"
Installation#
Docker (Recommended): Supports amd64/arm64. Key configs: WANWU_ARCH, WANWU_EXTERNAL_IP, WANWU_BFF_JWT_SIGNING_KEY. Access at http://localhost:8081 with default credentials admin/Wanwu123456.
Source Development: Use Makefile commands for individual service management on top of Docker deployment.
Roadmap#
Multimodal model integration, multimodal file parsing, API/database knowledge base import, general-purpose agents, A2A protocol, multi-agent systems, agent & model evaluation, monitoring & tracing, model playground, prompt engineering.