项目简介

A DeepResearch agent built with Spring AI Alibaba that conducts thorough research and writes polished reports. Using an LLM to call tools in a loop is the simplest form of an agent, but this can yield agents that are "shallow". DeepResearch implements task planning, sub agents, file system access, and a detailed prompt to solve complex, multi-step research tasks.

Quick Start

Clone Project

git clone https://github.com/alibaba/spring-ai-alibaba.git

cd spring-ai-alibaba/examples/deepresearch

Prerequisites

• JDK 17+
• AI_DASHSCOPE_API_KEY - Get from Alibaba Cloud Bailian
• JINA_API_KEY (Optional) - Get from Jina AI for MCP tool support

export AI_DASHSCOPE_API_KEY=your_dashscope_api_key
export JINA_API_KEY=your_jina_api_key  # Optional

Run the Agent

1. Run with Maven

mvn spring-boot:run

2. Or, Run in IDE Import as maven project and run Application.java.

Chat With Your Agent

After the agent starts, open in your browser: http://localhost:8080/chatui/index.html

Now you can ask the agent complex research questions like:

• "Research the core architecture design of Spring AI Alibaba"
• "Compare the technical implementations of Claude Code and GitHub Copilot"

Architecture

DeepResearch implements an agentic architecture that goes beyond simple tool-calling loops. It combines planning, context management, and subagent collaboration to tackle complex research tasks that require multiple steps and deep investigation.

Core Capabilities

🎯 Planning & Task Decomposition

• Uses TodoListInterceptor to break down complex tasks into discrete steps
• Dynamically tracks progress and adapts plans as new information emerges
• Supports parallel execution of independent research subtasks

📁 File System & Context Management

• Provides ls, read_file, write_file, edit_file, glob, grep tools for file operations
• Automatically offloads large tool results to filesystem to prevent context window overflow
• Enables persistence of research process and results across the conversation

🤝 Subagent Collaboration

• research-agent: Specialized for in-depth research, handles parallel sub-questions
• critique-agent: Reviews and improves report quality with detailed feedback
• Main agent coordinates subagents while keeping context clean and isolated

🔧 MCP Tool Integration

• Integrates external tools via Spring AI MCP Client
• Supports third-party services like Jina AI search
• Extensible to any MCP-compatible tools

Features provided by Spring AI Alibaba

Agent Framework

• ReactAgent: ReAct pattern implementation with reasoning-action loops
• Interceptor: Middleware mechanism for planning, file management, context editing
• Hook: Event hooks for human approval, summarization, tool call limits
• MemorySaver: In-memory state management with conversation history

Builtin Interceptors

• TodoListInterceptor: Task list management and tracking
• FilesystemInterceptor: File system access control with sandboxing
• SubAgentInterceptor: Subagent spawning and coordination
• ContextEditingInterceptor: Auto-compress context (triggers at 10K tokens)
• LargeResultEvictionInterceptor: Auto-dump large results to files
• ToolRetryInterceptor: Retry logic for failed tool calls

Builtin Hooks

• SummarizationHook: Conversation history summarization (triggers at 120K tokens)
• HumanInTheLoopHook: Human approval for critical operations
• ToolCallLimitHook: Limit tool calls per run (default: 25)

Integration Capabilities

• DashScope Chat Model support
• MCP Client for tool integration
• Agent Studio UI for visualization