deepsearch

DeepSearch

Overview

Deep Research Agent System Community Edition is a workflow-based deep research system capable of performing multi-round iterative web searches and comprehensive analysis on complex topics. The system combines LLM intelligent analysis with real-time web search to generate detailed research reports.

Dependent Skills

• byted-web-search

Workflow Invocation Logic

System Architecture

This system is a workflow description document with no execution script. It depends on the following external skills:

• byted-web-search skill: Execute web search`
• LLM Skill: Used for intelligent analysis and reasoning

Workflow

1. Input Reception Stage

• User Input: Research topic (sys.query) and research depth (depth)
• Default Depth: 3 (configurable)

2. Initialization Stage

• Create an iteration array with a depth equal to the specified depth value
• Array format:[0, 1, ..., depth-1]

3. Iterative Search Stage (executed over multiple rounds)

Each round of iteration executes the following steps:

a) LLM Intelligent Analysis

• Use the LLM to analyze the current research state
• Input: user query, collected findings, searched topics
• Output: JSON format, containing:
  • nextSearchTopic: The next search topic (string or None)
  • shouldContinue: Whether to continue searching (boolean value)

b) JSON Parsing

• ExtractnextSearchTopicandshouldContinuefields
• Update conversation variables

c) Conditional Judgment

• IfshouldContinueis True:
  • Execute byted-web-search:python scripts/web_search.py "<nextSearchTopic>"
  • Append the search results to the findings array
  • Update the search progress display
  • Continue to the next round of iteration
• IfshouldContinueis False:
  • End the current iteration
  • Output intermediate results

d) Variable Management

• UpdatenextSearchTopicandshouldContinuevariables
• AppendnextSearchTopicto thetopicsarray (records searched topics)
• Avoid repeatedly searching the same topic

4. Comprehensive Analysis Stage

• After all iterations are completed, use the LLM to comprehensively analyze all collected findings
• Generate a detailed comprehensive analysis report
• Output format: a detailed report in Markdown format

5. Report Generation Stage

• Output the final research analysis results
• Include important insights, conclusions, and remaining uncertainties
• Cite sources appropriately

Variable Description

The system maintains the following conversation variables:

Variable Name	Type	Description
topics	array[string]	List of searched topics
nextSearchTopic	string	The next topic to search
findings	array[string]	List of collected search results
shouldContinue	string	Flag indicating whether to continue searching

Web Search Integration

When web search is needed:

1. UsenextSearchTopicas the query parameter
2. Run the byted-web-search skill:python scripts/web_search.py "<query>"
3. Organize the answer based on the returned summary list, without adding or fabricating content
4. Append the search results to thefindingsarray

Note: Do not use any search parameter configurations (such as search_depth, topic, max_results, country, time_range, days, etc.); keep only the core input query.

Progress Tracking

The system displays search progress in real time:

• Format:{index + 1}/{depth}th search executed.
• For example:1/3th search executed.

Use Cases

Applicable Scenarios

1. Complex topic research: Requires in-depth, comprehensive research on a specific topic
2. Latest information analysis: Requires generating a detailed analysis report based on the latest web information
3. Multi-perspective exploration: Requires exploring a topic from different angles and dimensions
4. Systematic investigation: Requires systematic investigation and evidence collection

Typical Use Cases

• Market trend analysis
• Technology development research
• Competitor analysis
• Academic literature review
• Product research

Technical Features

Intelligent Features

1. Adaptive search: After each round of search, the LLM analyzes the results and intelligently decides the next search direction
2. Avoid repetition: The system records searched topics to avoid repeatedly searching the same content
3. Deep reasoning: Uses a dedicated reasoning model for comprehensive analysis

System Features

1. Multi-round iteration: Supports multi-round search with a specified depth
2. Parallel capability: Supports up to 10 parallel searches
3. State management: Complete variable management and state tracking
4. Visible progress: Displays search progress and status in real time

Integration Features

1. LLM integration: Combines GPT-4o for intelligent analysis and deepseek-reasoner for deep reasoning
2. Web search: Integrates the byted-web-search skill to obtain real-time web information
3. JSON processing: Uses JSON parsing tools to process structured data

Workflow Example

Input Example

用户查询: "人工智能在医疗领域的最新发展"
研究深度: 3

Execution Flow

1. Round 1:

   • LLM analysis: Decides to search "latest advances in AI medical diagnosis"
   • Web search: Executes the search and collects results
   • Status update: Records the topic and decides to continue searching

2. Round 2:

   • LLM analysis: Based on the results of Round 1, decides to search "breakthroughs in medical imaging AI technology"
   • Web search: Executes the search and collects results
   • Status update: Records the topic and decides to continue searching

3. Round 3:

   • LLM analysis: Based on the results of the first two rounds, decides to search "applications of AI in drug development"
   • Web search: Executes the search and collects results
   • Status update: Records the topic and decides to end the search

4. Comprehensive analysis:

   • LLM comprehensively analyzes all collected findings
   • Generate a detailed report on "The Latest Developments of Artificial Intelligence in the Healthcare Field"

Output Example

# 人工智能在医疗领域的最新发展研究报告

## 执行摘要
[基于三轮搜索的综合分析...]

## 主要发现
1. AI在医疗诊断方面的最新进展
   - [具体发现1]
   - [具体发现2]

2. 医疗影像AI技术突破
   - [具体发现3]
   - [具体发现4]

3. AI在药物研发中的应用
   - [具体发现5]
   - [具体发现6]

## 结论与建议
[综合分析结论...]

## 未来研究方向
[基于研究发现提出的未来研究方向...]

Precautions

Workflow Limitations

1. No executable script: This skill is a workflow description document and does not contain executable scripts
2. External dependencies: Depends on the byted-web-search skill to perform actual searches
3. Parameter simplification: When searching, only use the query parameter and ignore other search configurations

Usage Recommendations

1. Depth setting: Set an appropriate depth value according to research complexity
2. Query optimization: Provide a clear and specific research topic
3. Result verification: Perform necessary fact-checking on the generated report

Best Practices

1. Progressive research: Start with a broad topic and gradually go deeper into specific directions
2. Multi-source verification: Combine information from multiple sources for cross-verification
3. Timely updates: For rapidly developing topics, it is recommended to re-research regularly

Troubleshooting

Frequently Asked Questions

1. No search results: Check whether the query is too specific or specialized, and try more general search terms
2. Iteration ends too early: Adjust the LLM's temperature parameter or provide more context
3. Duplicate results: The system has a built-in mechanism to avoid duplication; if it still occurs, manual intervention can be used

Performance Optimization

1. Parallel search: Make full use of the system's parallel capabilities (up to 10 in parallel)
2. Cache utilization: For multiple research runs on the same topic, consider caching results
3. Incremental updates: For continuous research, an incremental update strategy can be adopted

---

Important Notice: This skill describes a deep research workflow, and actual execution depends on external skills and configurations. Please ensure that the byted-web-search skill and related LLM services are correctly configured.