参考资料
Mastra 官方资料
- Mastra GitHub
- Mastra Quickstart
- Mastra Project structure
- Mastra Build with AI
- Mastra Studio
- Model Providers
- Embedding models
- Gateway providers
- Agents overview
- Tools
- Tool streaming
- Structured output
- Processors
- Guardrails
- Agent approval
- Supervisor agents
- Background tasks
- Response caching
- Agent networks deprecated page
- Agent reference
- Agent.generate reference
- createTool reference
- Workflows overview
- Workflow control flow
- Workflow state
- Workflow snapshots
- Workflow suspend and resume
- Workflow agents and tools
- Workflow human-in-the-loop
- Workflow error handling
- Memory overview
- Storage overview
- Working memory
- Semantic recall
- Memory processors
- RAG overview
- RAG chunking and embedding
- RAG vector databases
- RAG retrieval
- MCP overview
- Request Context
- Streaming overview
- Streaming events
- Workflow streaming
- Deploy a Mastra server
- Middleware
- Auth overview
- Custom API routes
- Mastra Client SDK
- Storage overview
- Deployment overview
- Observability overview
- Evals / Scorers overview
- Custom scorers
- Running scorers in CI
- Mastra CLI reference
VitePress 官方资料
本教程采用的关键事实
| 事实 | 来源 |
|---|---|
mastra dev 会启动 Studio 和 REST endpoints | Mastra CLI reference |
src/mastra/index.ts 是 Mastra 相关代码入口 | Project structure |
模型可以用 provider/model 字符串表示,并可按请求动态选择 | Model Providers、Request Context |
| 模型 fallback 解决可用性问题,不等于质量提升 | Model Providers |
| embedding 模型也走模型路由,并要关注向量维度 | Embedding models、RAG vector databases |
| Agent 适合开放式任务,Workflow 适合确定步骤 | Agents overview、Workflows overview |
Tool 使用 createTool() 和 schema 定义输入输出 | createTool reference |
toModelOutput 用于控制工具结果给模型看到的形状,transform 用于 UI / transcript 展示 | Tools |
可以用 toolChoice 和 activeTools 控制运行时工具选择 | Tools |
| Memory 需要 resource/thread 绑定对话 | Memory overview |
| Working memory 可按 resource 或 thread 保存长期信息 | Working memory |
| Semantic recall 通过向量检索找回相似历史消息 | Semantic recall |
| RAG 的基本链路是 chunk、embedding、vector store、query | RAG overview |
| RAG chunk 策略要按文档类型选择,向量维度要匹配 embedding 模型 | RAG chunking and embedding、RAG vector databases |
| RAG 可用 rerank 提高检索结果相关性 | RAG retrieval |
| Workflow state 是跨步骤共享状态,普通业务数据仍应优先走 step input/output | Workflow state |
| Workflow suspend/resume 依赖 snapshot,snapshot 持久化需要 storage | Workflow snapshots、Workflow suspend and resume |
Structured output 通过 schema 约束 response.object | Structured output |
| RequestContext 用于请求级变量,不等同于长期 Memory | Request Context |
Streaming 可通过 textStream 和事件流观察 Agent 与 Workflow 执行 | Streaming overview、Streaming events |
| Processors 在模型调用前后转换、校验或控制消息 | Processors、Guardrails |
| Agent Approval 可以在工具执行前暂停并等待人工确认 | Agent approval |
| Supervisor Agent 通过子 Agent description 做委派,旧 Agent networks 页面已提示 deprecated | Supervisor agents、Agent networks deprecated page |
| Background tasks 适合长耗时工具调用,需要配置 storage | Background tasks |
| Response caching 当前是 alpha,适合重复请求,不适合带外部副作用的工具调用 | Response caching |
| Mastra Server 会把注册的 agents 和 workflows 暴露为 HTTP API,并支持 middleware 和 custom routes | Deploy a Mastra server、Middleware、Custom API routes |
| 鉴权后应把 resource ID 绑定到服务端认证结果,避免用户越权访问 thread | Middleware |
| 如果不配置 auth,Studio UI 和 API routes 会公开访问 | Auth overview |
| Storage 管理 suspended workflows、memory、traces 和 eval datasets | Storage overview |
| Observability 关注 trace、log、metric | Observability overview |
| Scorers 用于量化 Agent 输出质量,可用于 Studio、运行时或 CI | Evals / Scorers overview、Running scorers in CI |
| 自定义 scorer 的管线包含 preprocess、analyze、generateScore、generateReason,其中 generateScore 必需 | Custom scorers |
VitePress config 位于 <root>/.vitepress/config.ts | VitePress Site Config |