Dubbo Admin AI Documentation¶
Dubbo Admin AI is a set of runtime and service capabilities that brings AI features into Dubbo Admin. It combines model inference, tool invocation, knowledge retrieval, session management, and streaming output into a deployable service. Its goal is not "just chatting", but giving the system actionable capabilities for operations, troubleshooting, and knowledge Q&A.
If this is your first time looking at the project, treat it as three layers:
- Access layer: HTTP API + SSE streaming output. It receives requests, manages sessions, and pushes intermediate results to the client.
- Orchestration layer: the Agent thinks, decides whether to call tools, merges results, and produces the final answer.
- Capability layer: Models, Tools, Memory, and RAG provide model access, tools, short-term memory, and knowledge retrieval.
flowchart LR
Client["Client / Frontend"] --> Server["Server<br/>HTTP + SSE"]
Server --> Agent["Agent<br/>Think -> Act -> Observe"]
Agent --> Models["Models<br/>LLM / Embedding"]
Agent --> Tools["Tools<br/>Internal / Mock / MCP"]
Agent --> Memory["Memory<br/>Session Short-Term Memory"]
Tools --> RAG["RAG<br/>Retrieval Augmentation"]
RAG --> KB["Knowledge Base / Document Index"]
Where to start¶
- Want to run the service first: start with Quick Start.
- Want to understand the API and detailed YAML parameters: start with the User Guide.
- Want to understand the architecture and code: start with the Developer Guide.
Documentation scope¶
- User-facing guidance: how to deploy, integrate, and troubleshoot the service.
- Developer-facing guidance: runtime lifecycle, component boundaries, Agent workflow, the RAG subsystem, the tool system, and configuration mechanics.
- Evolution-oriented notes: why the system is designed this way, what the current constraints are, and what may change next.
Key facts¶
- The service listens on
http://localhost:8880by default. - The API base path prefix is
/api/v1/ai. - The streaming endpoint uses
text/event-stream. - Startup loads component config from
config.yaml, then runtime creates and initializes components in factory registration order. - The default component order is
logger -> memory -> models -> rag -> tools -> server -> agent.
Configuration overview¶
If you only want a high-level picture first, do not try to read every YAML file immediately. Focus on these files first:
config.yaml: the top-level assembly entry that decides which component config files are loaded.component/models/models.yaml: controls the default model, embedding model, and provider credentials.component/server/server.yaml: controls listen address, port, and timeouts.component/agent/agent.yaml: controls the model, prompt paths, and max iterations used by the Agent.component/tools/tools.yaml: controls whether mock, internal, and MCP tools are enabled.component/rag/rag.yaml: controls embedding, splitting, indexing, retrieval, and reranking for the knowledge pipeline.
For field-by-field explanations and full YAML semantics, go straight to YAML Configuration or the Configuration Guide.
Documentation structure¶
- Home: helps you build a fast mental model of the project.
- Quick Start: the shortest path to boot the service and send the first request.
- User Guide: focused on integration, deployment, and operations.
- Developer Guide: focused on architecture, maintenance, and capability extension.
- Proposals: capture major design decisions, constraints, and evolution plans.