wanwu

Yuanjing Wanwu (元景万悟) is an enterprise-oriented, multi-tenant AI agent development platform released by China Unicom under the Apache-2.0 license. It is positioned as a full-stack toolchain for Forward Deployed Engineers, bundling the capabilities needed to move an AI project from prototype to on-site delivery. Core capability areas described in the README: - RAG / knowledge base agent: end-to-end knowledge management covering creation, document parsing, vectorization, retrieval and re-ranking. Claims support for 12 file formats plus URL crawling, OCR and MinerU private deployment, multimodal retrieval, cascaded/adaptive chunking, intelligent ranking and source citation. A built-in UniAI-GraphRAG module adds domain ontology modeling for cross-document summarization and multi-hop reasoning. Knowledge bases created in Dify can reportedly be imported via API. - Ontology agent: builds business knowledge networks from enterprise data and documents to support structured-data reasoning and decision-making. - Workflow agent: a low-code drag-and-drop canvas with nodes for conditional branches, APIs, LLMs, knowledge bases, code and MCP, plus end-to-end debugging and performance analysis. The README also describes zero-code Skill invocation during agent development. - GUI agent: operates application interfaces without APIs, with isolated Docker sandbox containers per bot; a separate client download is linked. - General agent plus skill development: a dual-engine design combining a general agent with vertical-scene skills, natural-language skill creation, and one-click conversion of platform apps into skills. Platform modules include a model hub for unified access and lifecycle management of proprietary and open-source models (GPT, Claude, Llama and others), OpenAI API compatibility, multiple inference backends such as vLLM and TGI, and self-hosted options; an MCP module with 100+ recommended interfaces; web search with multi-source integration; visual workflow studio; high-precision RAG; the general agent and skills orchestration framework; the ontology agent; and Backend-as-a-Service RESTful APIs for integration with OA/CRM/ERP systems with fine-grained permission control. Deployment options are an out-of-the-box visual platform, API integration, and a Skill plus UniClaw dedicated client for high-privilege scenarios. The README lists multi-tenant architecture, Xinchuang compliance (Kunpeng CPUs, openEuler, Kylin, CULinux, TiDB, OceanBase) and an Apache-2.0 license as infrastructure and ecosystem features. Quick start is Docker-based: copy .env.example to .env, set WANWU_ARCH, WANWU_EXTERNAL_IP and WANWU_BFF_JWT_SIGNING_KEY, copy .env.ontology.example, optionally generate RSA keys, create a Docker network, then run docker compose with the amd64 or arm64 image env file. The default login is admin / Wanwu123456 at http://localhost:8081. Recommended hardware is 8-16 CPU cores, 32GB RAM, 200GB+ storage and no GPU; models used for the general agent or one-command skill creation need a context length of at least 32000. Source-code development, version upgrade, and TiDB/OceanBase adaptation procedures are also documented. Security notes mention RSA-encrypted password transmission and advise changing default passwords before production use. Typical scenarios cited are intelligent customer service, enterprise knowledge management, and process automation such as contract review and reimbursement approval. The README includes a feature comparison table against Dify.AI, FastGPT, RAGFlow and the Coze open-source version, dated May 15, 2026; these are the project's own claims and are not independently verified here.