DeepSearchAgents

VIBE 🖖 Build with 💖 for Humanity with AI

🚅

Come From Open Source, This is the Way

中文版

README Update Date: 2025-08-10

[v0.3.3.dev] Dev Status: "✅ Complete the development and integration of the DSCA front-end Alpha version"

1. Introduction

🚀 CODE ACT IS ALL YOU NEED

The DeepSearchAgent project embodies the philosophy that executable code as action is the most powerful paradigm for AI agents. By treating code generation and execution as the primary means of interaction with the world, we unlock unprecedented flexibility and capability in autonomous systems.

The DeepSearchAgent project is an intelligent agent system based on the ReAct (Reasoning + Acting) reasoning and action framework and the CodeAct ("Code as Action") AI agent concept. It aims to realize broader task reasoning and execution through "DeepResearch" DR-Multi-Agent, leveraging DeepSearch's multi-step network deep search foundational capabilities. It utilizes the reasoning power of AI language models (LLMs), along with a toolbox collection and programming action invocation abilities within a Python packages sandbox, enabling it to handle complex web search tasks that are both broad and deep via multi-step deep search, multimodal webpage text processing, reading, and multi-step reasoning. The system also provides traceable reference materials. Built upon Hugging Face's smolagents framework, this project implements a dual-mode intelligent agent system capable of invoking predefined toolboxes as well as writing action code—realizing both "generation of dedicated dynamic DSLs based on task plans" and "AI self-created dynamic one-time dedicated tools.

The project supports a CLI TUI interface for terminal command-line operation, a standard FastAPI service, the FastMCP MCP server, and a modern WebTUI frontend with terminal-style aesthetics. The WebTUI (v0.3.3) features real-time WebSocket streaming, a cyberpunk-inspired design system, and optimized performance for displaying the CodeAct Agent Run process. It facilitates developers in experimentation, integration, and usage across various systems. This is an open-source project aimed at providing VIBER beginners with a friendly Code Agent experience, learning opportunities, and extensibility.

2. ✨ Features

👻 Deep Search Task Capability: Handles complex questions through multi-step searching, reading, and reasoning processes involving online content.
DeepSearch Specialist: Supports both CodeAct (Python code execution) mode and ReAct (tool invocation) mode for experimental comparison; configuration of related Agent runtime, language models, and various tools can be managed via config.toml (src/core/config/settings.py).
🪄 Extensible Toolbox: Built-in set of tools for web searching, content retrieval, text processing, semantic ranking, computation, and GitHub repository analysis.
🌐 Hybrid Search Engine (v0.3.1): Multi-provider search aggregation supporting Google (Serper), X.com, Jina AI, and Exa Neural search with intelligent deduplication and ranking.
🌐 Web API v2 with Real-time WebSocket Streaming (v0.3.2): Simplified Gradio message pass-through architecture for web frontend integration with real-time agent execution visibility.
🔍 Text Embedding and Re-ranking: Uses Jina AI embedding and re-ranking models to process multimodal web content from URLs.
📚 GitHub Repository Q&A (v0.3.1): AI-powered repository analysis tool using DeepWiki MCP for understanding GitHub projects.
🐦 X.com Deep Query (v0.3.1): Specialized tools for searching, reading, and analyzing X.com (Twitter) content using xAI's Live Search API.
🧠 Periodic Planning and Updates: Implements strategic reevaluation during execution to optimize search paths.
🔄 Iterative Optimization: The AI specialist continuously improves search and analysis strategies based on initial findings through self-optimization, and continuously optimizes task execution paths by updating the task plan to achieve task objectives.
💻 Multiple development debugging and user interaction modes: Provides CLI command-line interaction, standard FastAPI service, and Web GUI frontend service
🔗 Traceable References: Provides sources and references for generated answers.
📺 Streaming Output: Supports real-time streaming of agent steps and final answers, with rich text formatting.
🧮 Symbolic Computation: Integrated WolframAlpha symbolic computation engine, supporting mathematical and computational problems
📝 JSON/Markdown Rendering: Automatically detects and presents structured outputs in user-friendly formats.
🤝 Hierarchical Multi-Agent Support (v0.2.9): Manager agent mode orchestrates teams of specialized agents for collaborative problem-solving.
⚡ Parallel Tool Execution (v0.2.9): Multiple concurrent tool calls for improved performance and efficiency.
📊 Enhanced Execution Metrics (v0.2.9): RunResult objects provide detailed execution metadata including token usage and timing.
🔒 Improved Security (v0.2.9): Latest security patches from smolagents v1.17.0-v1.19.0 applied.
🧠 Structured Generation (v0.2.9): Optional structured outputs for CodeAgent improving reliability.
🔄 Context Manager Support (v0.2.9): Proper resource cleanup lifecycle for better memory management.
💾 Enhanced Memory Management (v0.2.9): Agent memory reset and summary capabilities for long-running sessions.

Reference Use Cases (To be updated v0.3.1+)

WebTUI Demo:

CodeAct Mode Example: Full CLI run showing multi-step deep search process.
- Start:
- FinalAnswer:

Development Plan Currently Under Intensive Iteration:

[DONE] Developed Web API v2 with real-time WebSocket streaming (v0.3.2) - Simplified Gradio message pass-through architecture replacing complex event-driven system (~5000 lines reduced to ~500 lines). Frontend development and Docker packaging pending;
[DONE] Added MCP Client/MCP tools HUB to DeepSearchAgents' DeepSearchToolbox, supporting MCP Tools configuration and invocation;
[DONE] Provided packaging of DeepSearchAgents as an MCP server, offering DeepSearchAgent MCP tools services;
[DONE] Supported multi-vertical search engine source aggregation (Google, X.com, Jina AI, Exa Neural) with hybrid search aggregation and intelligent result deduplication (v0.3.1);
[DONE] Upgraded to smolagents v1.19.0 with hierarchical agent management, parallel tool execution, and enhanced streaming architecture;
[DONE] Add a DeepWiki Remote MCP tool to enhance the GitHub URLs vertical crawler/parser with GitHub Repository Q&A capabilities (v0.3.1);
[Partially supported in the tool layer] The deep search strategy provides more strategy parameters and adds support for strategy parameters based on Tokens budget;
[Experimental version testing] Implement auxiliary methods and tools for Agent Action search width & depth based on Monte Carlo Tree Search strategies in DeepSearchAgents, along with strategy control parameters;
[TODO] LLM as Judge: Experimentally add an Agent Runs evaluator for DeepSearchAgents (independently evaluate the deep search paths & results of DeepSearchAgents);
[TODO] Add persistent memory layer functionality for Agents & provide users with persistent search records;
Add suitable open-source sandbox (E2B-like) adapted code_sandbox Docker automation configuration, and increase support for more remote code_sandbox secure environment SDKs;
Integrate full-process Agent Runs telemetry adaptation ("OpenTelemetry" & Langfuse) (integrated together with the Docker packaged version);
[TODO] Human-in-the-loop & multi-path branching backtracking functionality for Agent Runs;
[Experimental] Special implementation version of multi_agent_HiRA (Hierarchical Reasoning Framework for Deep Search) based on special tokens protocol (arXiv-2507.02652v1);
[Experimental] Add auxiliary method optimization to the agent omni-tools-query pipeline based on [submodular-optimization] ("submodular optimization algorithm") to improve reQuery query effectiveness when using various external query tools; this auxiliary pipeline uses "submodular optimization algorithms" to optimize query selection, generate diversified tool search query inputs, perform effect evaluation, and return ReAct Agent action callbacks to help the Agent observe query result optimization effects in order to continuously iterate towards retrieval goals in subsequent Steps Actions. (https://jina.ai/news/submodular-optimization-for-diverse-query-generation-in-deepresearch/)

3. 🚀 Quick Start (CLI & FastAPI)

This section guides you through setting up the environment, installing dependencies, and running DeepSearchAgent via its command-line interface or standard FastAPI service.

Installation & Setup

Prerequisites:
- Python 3.13+.
- uv (Recommended, faster alternative to pip/venv): Install uv.
- Git.

Clone the Repository:

git clone https://github.com/DeepSearch-AgentTeam/DeepSearchAgent.git
cd DeepSearchAgent

Create Virtual Environment (Recommended):

# Using uv
uv venv
source .venv/bin/activate  # Unix/macOS
# .venv\Scripts\activate   # Windows

# Or using standard venv
# python -m venv .venv
# source .venv/bin/activate  # Unix/macOS
# .venv\Scripts\activate   # Windows

Install Dependencies:

For running the FastAPI service:
```
uv pip install .
```

For running the CLI or for Development:

# Installs core + CLI dependencies + development tools in editable mode
uv pip install -e ".[cli]"

For Development:
```
uv pip install -e ".[dev,test,cli]"
```

Configure:

# Create configuration files from templates
cp config.template.toml config.toml
cp .env.template .env

# Edit config.toml for models, agent parameters, service settings
# nano config.toml

# Edit .env for API keys (LITELLM_MASTER_KEY, SERPER_API_KEY, etc.)
# nano .env

Configuration Details:

config.toml: Contains non-sensitive configurations like model IDs, agent parameters (max steps, executor type), service settings (host, port).
.env: Contains only sensitive API keys (e.g., LITELLM_MASTER_KEY, SERPER_API_KEY, JINA_API_KEY, WOLFRAM_ALPHA_APP_ID).

(1) Running the CLI console

Ensure you have installed the CLI dependencies (see Step 4 in Installation & Setup).

# Run the CLI (interactive mode, uses settings from config.toml)
make cli
# or directly:
uv run python -m src.agents.cli

# Run with specific agent mode
python -m src.cli --agent-type react    # ReAct DeepSearch agent mode (Baseline ToolCalling mode)
python -m src.cli --agent-type codact   # CodeAct DeepSearch agent mode
python -m src.cli --agent-type manager  # Manager multi-agent mode (v0.2.9)

# Manager multi-agent mode with research team
python -m src.cli --agent-type manager --team research

CLI arguments will override settings defined in config.toml.

(2) Running the FastAPI Service

Ensure you have installed the core dependencies (see Step 4 in Installation & Setup).

# Start the main API server (uses host/port from config.toml, e.g., http://0.0.0.0:8000)
make run
# or directly:
uv run -- uvicorn src.agents.main:app --reload
# Note: --host and --port are now taken from config.toml via main.py
# Use LOG_LEVEL environment variable for log level (e.g., LOG_LEVEL=debug make run)

API Endpoints:

POST /run_codact_agent: Runs the CodeAct DeepSearch agent.
POST /run_deepsearch_agent: Runs the agent configured by service.deepsearch_agent_mode in config.toml (or DEEPSEARCH_AGENT_MODE env var).
GET /: API info and health check.

Example API request to the configured deep search REST API endpoint:

curl -X POST http://localhost:8000/run_deepsearch_agent \
  -H "Content-Type: application/json" \
  -d '{"user_input": "Search the latest news about OpenAI'''s new GPT-4.1 API."}'

(Replace localhost:8000 with the actual host and port if changed in config.toml)

(3) Using the Web API v2 with WebTUI Frontend

The Web API v2 provides real-time WebSocket streaming for the WebTUI frontend integration. This major refactoring (v0.3.2-v0.3.3) delivers a clean, maintainable architecture that processes agent events directly through the web_ui.py module, paired with a modern terminal-style web interface.

Architecture Overview:

Agent (React/CodeAct) → stream_agent_messages() → DSAgentMessageProcessor → WebSocket → Frontend

WebTUI Frontend Features:

Terminal-Style Interface: Cyberpunk-inspired design with terminal aesthetics
Real-time Streaming: Watch agent reasoning, tool execution, and results as they happen
Design System Components: Custom DS components for consistent terminal-style UI
Performance Optimized: 40% CSS reduction, 80% animation CPU reduction (v0.3.3)
Three Display Modes:
- Chat Interface: Message display with thought truncation and markdown rendering
- Code Editor: Monaco-based syntax-highlighted code viewing
- Terminal Output: Real-time execution logs and results

Backend WebAPI WebSocket Event:

Direct Event Processing: Handles 4 main event types from DeepSearchAgent WebAPI:
- PlanningStep: Agent planning with strategy updates
- ActionStep: Tool execution with thoughts and results
- FinalAnswerStep: Structured final answers
- ChatMessageStreamDelta: Real-time streaming updates
Metadata-Driven Routing: Frontend components selected via message metadata
Session Management: Multi-turn conversations with session isolation
Component Routing: Messages routed to chat, webide, or terminal based on content

Quick Start:

// Connect to WebSocket
const ws = new WebSocket('ws://localhost:8000/api/v2/ws/my-session?agent_type=codact');

// Handle incoming messages with component routing
ws.onmessage = (event) => {
  const message = JSON.parse(event.data);
  
  // Route based on metadata
  if (message.metadata?.component === 'chat') {
    // Display in chat (planning, thoughts, final answers)
  } else if (message.metadata?.component === 'webide') {
    // Show in code editor (Python execution)
  } else if (message.metadata?.component === 'terminal') {
    // Display in terminal (tool outputs, logs)
  }
};

// Send query
ws.send(JSON.stringify({type: 'query', query: 'Your question here'}));

Message Format:

Each DSAgentRunMessage includes routing metadata
Message types: planning_header, action_thought, tool_call, final_answer, etc.
Streaming support with initial → delta → final pattern

See src/api/v2/README.md for comprehensive documentation, src/api/v2/STREAM_EVENTS.md for event flow details, and src/api/v2/examples/ for example implementations.

Starting the WebTUI Frontend:

# Navigate to frontend directory
cd frontend

# Install dependencies
npm install

# Start development server
npm run dev

# The WebTUI will be available at http://localhost:3000

For production deployment:

npm run build
npm start

See frontend/README.md for detailed frontend documentation and configuration options.

(4) Running the MCP Server (MCP Tools `deepsearch_tool`)

DeepSearchAgent now supports serving as a Model Context Protocol (MCP) server, exposing deep search capabilities as an MCP tool deepsearch_tool that can be accessed by any MCP client.

# Run the FastMCP server with default settings
python -m src.agents.servers.run_fastmcp
# or
python -m src.agents.servers.run_fastmcp --agent-type codact --port 8100

This starts a FastMCP server with Streamable HTTP transport at http://localhost:8100/mcp (default), providing access to DeepSearchAgent's functionality through the deepsearch_tool endpoint.

Server Arguments:

--agent-type: Agent type to use (codact or react, default: codact)
--port: Port number for server (default: 8100)
--host: Host address (default: 0.0.0.0)
--debug: Enable debug logging
--path: Custom URL path (default: /mcp)

Debugging with MCP Inspector:

The MCP Inspector can be used to debug and interact with the DeepSearchAgent MCP server:

Install MCP Inspector if you haven't already: It is also recommended to consider using the forked enhanced version of the MCP Inspector client development debug console: MCPJam-Inspector

npm install -g @modelcontextprotocol/inspector

Start the MCP Inspector dev MCP Client debugging console:

npx @modelcontextprotocol/inspector

In the browser UI that opens (typically at http://127.0.0.1:6274):
- Set Transport Type: Streamable HTTP
- Set URL: http://localhost:8100/mcp
- Click "Connect"
- Navigate to "Tools" tab and select "deepsearch_tool"
- Enter your search query and click "Run Tool" button
You'll see real-time progress updates and the final search results rendered in the MCP Inspector Web UI.

FastMCP Server in FastAPI Application:

You can also embed the FastMCP server in the main FastAPI application:

# Run main API server with FastMCP integration
python -m src.main --enable-fastmcp --agent-type codact

When run with --enable-fastmcp, the main API server mounts the FastMCP server at /mcp-server (default) for integrated operation.

(5) Frontend Development Quick Start

DeepSearchAgents includes a modern web frontend built with Next.js that provides a rich interface for interacting with agents through the Web API v2.

Technology Stack:

Framework: Next.js 15.3.3 with App Router
UI Library: React 19 with TypeScript
Styling: TailwindCSS with custom design tokens
Components: Radix UI primitives, Monaco Editor, XTerm.js
Real-time: WebSocket with automatic reconnection
Rendering: React Markdown with syntax highlighting

Setup Frontend Development Environment:

# Navigate to frontend directory
cd frontend

# Install dependencies
npm install
# or
yarn install

# Create environment variables
cp .env.example .env.local

# Edit .env.local to set backend URLs
# NEXT_PUBLIC_API_URL=http://localhost:8000
# NEXT_PUBLIC_WS_URL=ws://localhost:8000

Development Commands:

# Start development server (http://localhost:3000)
npm run dev

# Build for production
npm run build

# Run production server
npm run start

# Run linting
npm run lint

Key Frontend Features:

Component-Based Routing: Messages automatically routed to appropriate UI components
Real-time Streaming: See agent thoughts, actions, and results as they happen
Rich UI Components:
- AgentChat: Main chat interface with message grouping
- ActionThoughtCard: Displays agent reasoning (truncated to 60 chars)
- PlanningCard: Shows planning steps with badges
- FinalAnswerDisplay: Structured final answer rendering
- Code editor for Python execution
- Terminal for tool outputs
Session Management: Persistent sessions with conversation history
Responsive Design: Optimized for desktop and mobile

Frontend Integration Example:

// Using the WebSocket hook
import { useWebSocket } from '@/hooks/use-websocket';

function MyComponent() {
  const { messages, sendMessage, isConnected } = useWebSocket({
    sessionId: 'my-session',
    agentType: 'codact'
  });

  const handleQuery = () => {
    sendMessage({
      type: 'query',
      query: 'Search for latest AI developments'
    });
  };

  return (
    <div>
      {messages.map(msg => (
        <AgentChat key={msg.message_id} message={msg} />
      ))}
    </div>
  );
}

See frontend/README-DeepSearchAgents.md for comprehensive frontend documentation and development guidelines.

4. 🛠️ Architecture and Modules

The core system architecture includes:

Core Specialist Agents Modules (src/agents/react_agent.py, src/agents/codact_agent.py, src/agents/manager_agent.py): Implement ReAct, CodeAct, and Manager agent logic based on smolagents. Manager agent (v0.2.9) orchestrates teams of specialized agents for collaborative problem-solving.
Specialist Core Agent Runtime Module (src/agents/runtime.py): Responsible for managing the agent's runtime environment, including hierarchical agent orchestration.
Agent Toolkit Collection (src/tools/): Functions that the agent can invoke (such as web search, reading URLs, etc.).
FastAPI Service (src/api): FastAPI service providing REST API related services.
CLI Interface (src/cli.py): Provides an interactive command-line interface with rich formatting.
Web API v2 (src/api/v2/): Real-time WebSocket API with direct agent event processing via web_ui.py and DSAgentMessageProcessor, transforming smolagents events into metadata-rich messages for frontend component routing.
MCP Server (src/agents/servers/run_fastmcp.py): FastMCP server providing MCP tools services with Streamable HTTP transport.

Web API v2 Architecture (v0.3.2.rc2)

The Web API v2 features a streamlined architecture that processes AI agent long-running multi-step task events directly:

Event Processing Pipeline:

smolagents Events → web_ui.py → DSAgentMessageProcessor → WebSocket → Frontend Components

Key Components:

web_ui.py: Core event processor handling 4 main event types from smolagents
DSAgentMessageProcessor: Wraps processed events into DSAgentRunMessage format
Session Manager: Handles multi-session support and agent lifecycle
WebSocket Endpoint: Streams messages with extensive metadata for frontend routing

Event Type Mapping:

PlanningStep → [InitialPlanning&UpdatePlanning] Multiple messages: planning_header, planning_content, planning_footer
ActionStep → [ActionReasoning&CodeGeneration&SandboxExecution&Observation] Multiple messages: action_header, action_thought, tool_call, execution_logs
FinalAnswerStep → [FinalAnswerStructuredOutput] final_answer with structured data
ChatMessageStreamDelta → Streaming deltas for real-time updates

Architecture diagram updated for version v0.3.2.rc2

See docs/architecture-diagram/architecture-diagram-v0.3.2.rc2.mmd for the latest architecture diagram details.

5. ⚙️ Agent Modes (ToolCalling ReAct vs CodeAct)

DeepSearchAgent supports two modes of agent operation: the CodeAct code-execution mode and the ReAct tool-calling mode. The default mode used by the /run_deepsearch_agent endpoint is configured in config.toml (service.deepsearch_agent_mode) or via the DEEPSEARCH_AGENT_MODE environment variable.

Both modes now support streaming output, providing real-time visibility into the agent's reasoning and execution process.

CodeAct Mode (Code Execution)

In CodeAct mode, the agent produces executable Python code, and by running that code the agent completes its reasoning and actions. This allows for more complex operations, combining multiple steps into one code execution.

Example format:

results = search_links("example query")
content = read_url(results[0]["link"])
final_answer("The result is...")

ReAct (Function Calling Tool) Baseline Mode

In ReAct mode, the agent operates in the classic reasoning+acting manner, with actions executed by invoking predefined tools. During its reasoning process, the LLM generates structured "action" outputs that specify which tool to use and with what parameters.

Example format:

{
  "name": "search_links",
  "arguments": {
    "query": "example query"
  }
}

Comparison and Use Cases

Differences	ToolCalling ReAct Mode	CodeAct Mode
Action Representation	Structured JSON instructions	Executable Python code
Complex Operation Capability	Multiple steps for complex logic	Can combine multiple steps with programming constructs
Model Requirements	General conversational ability	Requires code generation capability
Debugging & Interpretability	Human-readable thoughts and actions	Code traces with error feedback
Best For	Simple queries, fixed workflows	Complex tasks, flexible tool orchestration
Streaming Support	Support	Support
Planning Capability	Periodic planning every N steps	Periodic planning every N steps

Manager Mode (Hierarchical Multi-Agent) - v0.2.9

Manager mode introduces hierarchical agent orchestration, where a manager agent coordinates a team of specialized agents to solve complex problems collaboratively. This mode leverages the managed agents support added in smolagents v1.19.0.

Architecture:

Manager Agent: High-level orchestrator that breaks down complex queries
Specialized Agents: Team members with specific expertise (ReAct or CodeAct agents)
Delegation Logic: Manager assigns subtasks to the most appropriate agent

Research Team Configuration:

The default research team includes:

Web Research Specialist (ReAct Agent): Focuses on web search, content retrieval, and information gathering
Data Analysis Specialist (CodeAct Agent): Handles data processing, computation, and synthesis

Example Usage:

# CLI with research team
python -m src.cli --agent-type manager --team research

# Custom team configuration
python -m src.cli --agent-type manager --team custom --managed-agents react codact

Benefits:

Collaborative Problem Solving: Different agents handle their specialty areas
Improved Accuracy: Combines strengths of different agent paradigms
Scalability: Easy to add new specialized agents to teams
Task Parallelization: Manager can delegate multiple subtasks concurrently

Feature	Manager Mode
Agent Coordination	Hierarchical delegation to specialized agents
Complex Query Handling	Breaks down into subtasks for team members
Model Requirements	Orchestration + specialized agent capabilities
Best For	Multi-faceted research, comparative analysis, complex workflows
Team Composition	Configurable teams of ReAct/CodeAct agents

6. 🔧 Agent Toolbox Chain

DeepSearchAgent comes with an extensible toolchain that helps the agent retrieve and process information. These tools work in concert to form a complete query-answering pipeline:

Search and Discovery Tools

search_links: Accepts a query string and uses external search engine APIs to retrieve web results with titles, snippets, and URLs. Enhanced in v0.3.1: Now supports hybrid search with multiple providers:
- Serper API (Google): Traditional web search for comprehensive coverage
- X.com (xAI API): Real-time social media content from X.com/Twitter with live data access
- Jina AI Search: LLM-optimized search with advanced content extraction
- Exa Neural Search: Semantic search with neural understanding
- Hybrid Aggregation: Intelligent deduplication and ranking across all providers
- Auto-detection: Automatically selects the appropriate search engine based on query content
search_fast (v0.3.1): Optimized search tool for speed-critical operations
github_repo_qa (v0.3.1): AI-powered GitHub repository analysis using DeepWiki MCP

Content Retrieval and Processing Tools

read_url: Fetches HTML content from standard web pages and extracts formatted text for analysis. Enhanced in v0.3.1 with modular scraping architecture:
- Auto-provider selection: Chooses best scraper based on URL and availability
- JinaReader: LLM-optimized content extraction
- Firecrawl: Advanced JavaScript rendering support
- Fallback mechanism: Automatic failover between providers
xcom_read_url: Specialized tool for reading X.com (Twitter) content using xAI's Live Search API. Provides real-time access to posts, profiles, and search results.
xcom_qa (v0.3.1): Deep Q&A tool for X.com content analysis with search, read, and query operations
chunk_text: Splits long text into manageable segments for detailed analysis using intelligent segmentation.
embed_texts: Encodes text chunks into vector representations for semantic similarity operations.
rerank_texts: Ranks text chunks by relevance to a given query for finding the most relevant information.

Calculation & Scientific Query Tools

wolfram (Computation Engine): Calls the WolframAlpha API to handle mathematical calculations or scientific queries.

Structured Output Tools

final_answer (Final Answer): Indicates that the agent has reached a conclusion, using structured output & terminating the reasoning loop.

Toolbox Management System

NEW in v0.2.8: The toolbox.py module provides a unified interface for managing DeepSearchAgent tools:

Tool Registry: Centralized registration system for all built-in and external tools
Factory Methods: Automated tool instantiation with proper API key configuration
Extension Support: Integration with Hugging Face Hub collections and MCP (Model Context Protocol) servers
Configuration Loading: Automatic tool loading based on config.toml settings

Toolbox Features:

# Create tool collection with specific tools
toolbox.create_tool_collection(
    api_keys=api_keys,
    tool_names=["search_links", "read_url", "xcom_read_url"],
    verbose=True
)

# Load tools from Hub collections
toolbox.load_from_hub("collection_slug", trust_remote_code=True)

# Load tools from MCP servers
with toolbox.load_from_mcp(server_params, trust_remote_code=True):
    # Use tools from MCP server
    pass

Enhanced Search Workflow Task Reasoning Prompting Template

In the typical v0.3.1 enhanced sequence:

Hybrid Enhanced Search: The agent specialist uses search_links, which automatically detects whether the query is related to X.com content (mentions @username, hashtags, trending topics) and routes it to the appropriate search engine.
Content Parsing: Depending on the source, use read_url to handle standard web content or xcom_read_url for X.com content.
Content Processing Pipeline: For complex content, use chunk_text, embed_texts, and rerank_texts to identify key paragraphs.
Computational Analysis: When calculations are needed, call wolfram for mathematical analysis.
Final Answer: This loop continues until the agent specialist determines there is enough information and calls final_answer.

Multi-Source Agent Specialist Collaborative Search

The enhanced toolchain now offers:

Hybrid Search Capability: Traditional web search, real-time social media, semantic search, and codebase analysis.
Adaptive Parsing: Different extraction strategies and pipeline parsing tools tailored for various content types.
Unified Interface: Consistent tool invocation regardless of underlying data sources.
Real-Time Data: Access to real-time social media content via xAI integration.
Code Repository Understanding: AI-driven code repository analysis through DeepWiki integration.

7. 💡 Theoretical Foundations

CodeAct Executable Code Agents | CodeAct Executable Code Agents

CodeAct refers to a method where agents generate and execute actions in the form of code. The core idea is that at each decision step, the model directly produces executable code, which is run to invoke tools or perform computations.

Compared to static instructions, using code as an action representation offers greater expressive power and flexibility: it can combine multiple tool calls, apply programming logic to handle complex data structures, and even reuse previously defined functions, greatly expanding the agent’s action space.

Periodic Planning and Adaptive Search | Periodic Planning and Adaptive Search

The agent mode implements periodic planning intervals, allowing the agent to reassess its strategy every N steps. This enables more efficient search paths by:

Evaluating progress relative to the original task
Identifying gaps in information gathering
Adjusting search direction based on discovered content
Prioritizing new search avenues when current approaches are ineffective

References

ReAct: Synergizing Reasoning and Acting in Language Models arXiv:2210.03629v3

Executable Code Actions Elicit Better LLM Agents arXiv:2402.01030v4

DynaSaur: Large Language Agents Beyond Predefined Actions arXiv:2411.01747v1

LLMCompiler: An LLM Compiler for Parallel Function Calling arXiv:2312.04511v3

ReWOO: Decoupling Reasoning from Observations for Efficient Augmented Language Models arXiv:2305.18323v1

smolagents.agents.CodeAgent

Jina AI DeepResearch repository

A Practical Guide to Implementing DeepSearch/DeepResearch

8. 📦 Installation

Requirements

Python 3.13+
Create config.toml from config.template.toml and customize parameters.
Create .env from .env.template and add required API keys:
- LITELLM_MASTER_KEY (if using LiteLLM compatible models)
- SERPER_API_KEY (for web search via search_links)
- JINA_API_KEY (for content processing via read_url, embed_texts, rerank_texts)
- WOLFRAM_ALPHA_APP_ID (optional, for computational queries via wolfram)
- LITELLM_BASE_URL (optional, if using a custom LiteLLM endpoint)
- etc. Other Agent Tools may require additional API keys.

9. 🤝 Contributing

Contributions are welcome! Please feel free to submit a Pull Request.

10. 📄 License

This project is licensed under the MIT License

11. 📝 Acknowledgements Open Source Projects

Special thanks to the following open-source projects (as well as other equally important projects not listed), "May the Force be with you":

12. 👨‍💻 AI Coder Pair Programming Assistance & 🖖 VIBE Best Practices Sharing for Programming

The DeepSearchAgent project is designed with modern AI engineers and human engineers collaborating in software development and coding workflows in mind. We have integrated special repository workspace rule files (.cursor/rules/*.mdc) to facilitate AI-assisted development and ensure consistency within the codebase.

Equivalent to `.cursor/rules/` (`.mdc`)

CLAUDE.md: Claude Code Prompting markdown file.
AGENTS.md: Codex CLI & Codex Software engineering Agent, Prompting markdown file.

Use `.cursor/rules/` Files

This repository contains special Markdown files in the .cursor/rules/ & CLAUDE.md directory, serving as contextual guideline prompts for human developers and AI coding assistants. These files are similar to the CLAUDE.md concept described in Claude Code Best Practices, providing structured information about project architecture, components, and conventions.

VIBE Programming & Development References:

claude-code-best-practices

vibe-coding-higher-quality-code

Coding Agents 101: The Art of Actually Getting Things Done

Available Rules Files

CLAUDE.md: Claude Code rules & memory file.
agent-architecture.mdc: Documents the agent design patterns (ReAct, CodeAct, and Manager) and their functionalities
configuration.mdc: Provides detailed explanations of the customized configuration system options
interfaces.mdc: Describes the available interfaces (CLI, FastAPI, MCP Tool Server)
jina-ai-api-rules.mdc: Contains guidelines for using various Jina AI APIs within the codebase
periodic-planning.mdc: Explains the cyclical planning features used for strategic reassessment
project-overview.mdc: Offers a comprehensive overview and structure of the project
tools.mdc: Records the functionalities of dedicated tools used for web search, content processing, and analysis

Benefits for Developers

These rules files facilitate collaboration between human developers and AI engineers by:

Quick Onboarding: Helping AI engineers rapidly understand the project architecture and design decisions
Consistent Development: Ensuring code adheres to established patterns and conventions
AI-Assisted Development: Providing context to AI engineers for generating more accurate and relevant code
Code as Documentation: Keeping documentation closely integrated with the code for easy access

Best Practices for AI Engineer Collaboration

When using AI to assist in developing this project, we recommend the following workflow:

Understand the Rules: Review the .cursor/rules/*.mdc files related to the component you are developing.
Reference Specific Rules: When collaborating with AI engineers, explicitly cite the relevant rule files.
Iterative Improvement: Use AI for initial code generation, then refine the solution according to project conventions.
Planning for Complex Changes: For complex features, have AI outline the plan before generating implementation code.
Test-Driven Approach!: For key components, work with AI engineers to write tests before implementing the code!
Update Rules: When making significant changes, update the relevant rule files.

Example Workflow

Exploring the Codebase

When pairing with AI engineers to explore the codebase, you can start like this:

Please help me understand the DeepSearchAgent architecture. Refer to .cursor/rules/project-overview.mdc and .cursor/rules/agent-architecture.mdc for details.

Add New Feature

When adding a new tool to the tool collection:

I need to add a new tool for YouTube video analysis. Please implement it following the pattern in .cursor/rules/tools.mdc and the code style in .cursor/rules/python-code-style-pep8.mdc.

Update Configuration

When modifying the configuration system:

I need to add new configuration options for depth search Tokens budget & index depth. Please advise on how to extend the configuration structure according to .cursor/rules/configuration.mdc.

Contributing to Rules

As the project evolves, we encourage contributors to update and expand these rules files. If you're adding a new major component or changing existing architecture, please update the relevant .mdc files to reflect these changes. This helps maintain the documentation as a living resource that accurately represents the current state of the codebase.

13. Project Structure

src/
├── agents/                   # Agent implementations and core logic
│   ├── prompt_templates/     # Modular prompt template system
│   │   ├── __init__.py
│   │   ├── codact_prompts.py # CodeAct agent prompts and templates
│   │   └── react_prompts.py  # ReAct agent prompts and templates
│   ├── servers/              # Server implementations
│   │   ├── __init__.py
│   │   ├── run_fastmcp.py    # FastMCP MCP server implementation
│   │   └── run_fastmcp.py    # FastMCP MCP server
│   ├── ui_common/            # Shared UI components and utilities
│   │   ├── __init__.py
│   │   ├── agent_step_callback.py     # Agent execution step callbacks
│   │   ├── console_formatter.py       # Console output formatting
│   │   ├── constants.py               # UI-related constants
│   │   └── streaming_formatter.py     # Streaming output formatter (v0.2.9)
│   ├── __init__.py
│   ├── base_agent.py         # Base agent interface and common functionality
│   ├── codact_agent.py       # CodeAct agent implementation
│   ├── manager_agent.py      # Manager agent implementation (v0.2.9)
│   ├── react_agent.py        # ReAct agent implementation
│   ├── run_result.py         # Agent run result objects (v0.2.9)
│   ├── runtime.py            # Agent runtime manager
│   └── stream_aggregator.py  # Stream aggregation logic (v0.2.9)
├── api/                      # FastAPI service components
│   ├── v1/                   # API version 1 implementation
│   │   ├── endpoints/        # API endpoint definitions
│   │   │   ├── __init__.py
│   │   │   ├── agent.py      # Agent-related endpoints
│   │   │   └── health.py     # Health check endpoints
│   │   ├── __init__.py
│   │   └── router.py         # API router configuration
│   ├── v2/                   # API version 2 (v0.3.2) - WebSocket streaming
│   │   ├── examples/         # Example client implementations
│   │   │   ├── test_agent_steps_full.js  # WebSocket streaming test
│   │   │   ├── test_debug.py             # Direct processor testing
│   │   │   └── test_simple_agent.py      # Agent integration example
│   │   ├── __init__.py
│   │   ├── endpoints.py      # WebSocket and REST endpoints
│   │   ├── web_ui.py         # Core event processing and routing
│   │   ├── ds_agent_message_processor.py  # Message processor wrapper
│   │   ├── main.py           # Standalone API server
│   │   ├── models.py         # Pydantic data models
│   │   ├── openapi.yaml      # OpenAPI specification
│   │   ├── README.md         # Comprehensive v2 documentation
│   │   ├── STREAM_EVENTS.md  # Event flow documentation
│   │   ├── session.py        # Session management
│   │   └── WebAPIv2-GUI-Interface-API-Docs.md  # Frontend integration guide
│   ├── __init__.py
│   └── api.py                # Main API configuration
├── core/                     # Core system components
│   ├── chunk/                # Text chunking components
│   │   └── segmenter.py      # Text Chunker
│   ├── config/               # Configuration handling
│   │   ├── __init__.py
│   │   └── settings.py       # Settings management and configuration loading
│   ├── ranking/              # Content ranking and embedding
│   │   ├── __init__.py
│   │   ├── base_ranker.py    # Base ranking interface
│   │   ├── chunker.py        # Text chunking utilities
│   │   ├── jina_embedder.py  # Jina AI Token Embedder
│   │   └── jina_reranker.py  # Jina AI Reranker
│   ├── scraping/             # Web content scraping (v0.3.1 refactored)
│   │   ├── __init__.py
│   │   ├── base.py           # Base scraper abstraction
│   │   ├── result.py         # Scraping result data structures
│   │   ├── scrape_url.py     # Main scraping orchestrator
│   │   ├── scraper_firecrawl.py  # Firecrawl Scraper
│   │   ├── scraper_jinareader.py # JinaReader Scraper
│   │   ├── scraper_xcom.py   # X.com (Twitter) specialized scraper
│   │   └── utils.py          # Scraping utility functions
│   ├── search_engines/       # Search engine integrations (v0.3.1 expanded)
│   │   ├── utils/            # Search utility modules
│   │   │   ├── __init__.py
│   │   │   └── search_token_counter.py  # Token counting utilities
│   │   ├── __init__.py
│   │   ├── base.py           # Base search client abstraction
│   │   ├── search_exa.py     # Exa Search
│   │   ├── search_hybrid.py  # Hybrid Search Aggregator
│   │   ├── search_jina.py    # Jina AI Search
│   │   ├── search_serper.py  # Serper API (Google) search engine
│   │   ├── search_xcom.py    # X.com search base
│   │   └── search_xcom_sdk.py # X.com SDK implementation
│   ├── github_toolkit/       # GitHub integration tools (v0.3.1)
│   │   ├── __init__.py
│   │   └── deepwiki.py       # DeepWiki Remote MCP client wrapper
│   ├── xcom_toolkit/         # X.com toolkit (v0.3.1)  
│   │   ├── __init__.py
│   │   └── xai_live_search.py # xAI Live Search client
│   └── __init__.py
├── tools/                    # Agent Tools Collection (v0.3.1 expanded)
│   ├── __init__.py
│   ├── chunk.py              # Text chunking tool
│   ├── embed.py              # Text embedding tool
│   ├── final_answer.py       # Final answer formatter tool
│   ├── github_qa.py          # GitHub repository Q&A tool (v0.3.1)
│   ├── readurl.py            # Generic URL content reading tool
│   ├── rerank.py             # Content reranking tool
│   ├── search.py             # Multi-engine Hybrid web search tool
│   ├── search_fast.py        # Fast search tool (v0.3.1)
│   ├── search_helpers.py     # Search helper utilities (v0.3.1)
│   ├── toolbox.py            # Tool management and registry system
│   ├── wolfram.py            # Symbolic computational tool
│   ├── xcom_qa.py            # X.com Deep Q&A tool (v0.3.1)
│   └── xcom_readurl.py       # X.com (Twitter) URL reading tool
├── cli.py                    # Command-line interface version
└── main.py                   # FastAPI main entry

frontend/                     # Next.js web frontend for DSCA WebTUI (v0.3.3)

Name		Name	Last commit message	Last commit date
Latest commit History 164 Commits
.claude		.claude
.cursor/rules		.cursor/rules
.gemini		.gemini
.github/workflows		.github/workflows
debug-archive		debug-archive
dev-docs		dev-docs
docs		docs
frontend		frontend
public		public
src		src
tests		tests
.coveragerc		.coveragerc
.env.example		.env.example
.gitignore		.gitignore
CHANGELOG.md		CHANGELOG.md
CLAUDE.md		CLAUDE.md
CRUSH.md		CRUSH.md
LICENSE		LICENSE
Makefile		Makefile
README.md		README.md
README_Zh.md		README_Zh.md
config.template.toml		config.template.toml
pyproject.toml		pyproject.toml
pytest.ini		pytest.ini
requirements.txt		requirements.txt

Folders and files

Latest commit

History

Repository files navigation

DeepSearchAgents

🚅

1. Introduction

🚀 CODE ACT IS ALL YOU NEED

2. ✨ Features

3. 🚀 Quick Start (CLI & FastAPI)

Installation & Setup

(1) Running the CLI console

(2) Running the FastAPI Service

(3) Using the Web API v2 with WebTUI Frontend

(4) Running the MCP Server (MCP Tools deepsearch_tool)

(5) Frontend Development Quick Start

4. 🛠️ Architecture and Modules

Web API v2 Architecture (v0.3.2.rc2)

5. ⚙️ Agent Modes (ToolCalling ReAct vs CodeAct)

CodeAct Mode (Code Execution)

ReAct (Function Calling Tool) Baseline Mode

Comparison and Use Cases

Manager Mode (Hierarchical Multi-Agent) - v0.2.9

6. 🔧 Agent Toolbox Chain

Search and Discovery Tools

Content Retrieval and Processing Tools

Calculation & Scientific Query Tools

Structured Output Tools

Toolbox Management System

Toolbox Features:

Enhanced Search Workflow Task Reasoning Prompting Template

Multi-Source Agent Specialist Collaborative Search

7. 💡 Theoretical Foundations

CodeAct Executable Code Agents | CodeAct Executable Code Agents

Periodic Planning and Adaptive Search | Periodic Planning and Adaptive Search

References

8. 📦 Installation

Requirements

9. 🤝 Contributing

10. 📄 License

11. 📝 Acknowledgements Open Source Projects

12. 👨‍💻 AI Coder Pair Programming Assistance & 🖖 VIBE Best Practices Sharing for Programming

Equivalent to .cursor/rules/ (.mdc)

Use .cursor/rules/ Files

Available Rules Files

Benefits for Developers

Best Practices for AI Engineer Collaboration

Example Workflow

Exploring the Codebase

Add New Feature

Update Configuration

Contributing to Rules

13. Project Structure

About

Resources

License

Uh oh!

Stars

Watchers

Forks

Releases 4

Packages 0

Uh oh!

Uh oh!

Contributors

Uh oh!

Languages

(4) Running the MCP Server (MCP Tools `deepsearch_tool`)

Equivalent to `.cursor/rules/` (`.mdc`)

Use `.cursor/rules/` Files

Packages