home / skills / lin-a1 / skills-agent / deepsearch_service

deepsearch_service skill

needs review

This skill performs iterative deep-search and reasoning to decompose complex queries, gather multi-source info, and generate structured reports.

npx playbooks add skill lin-a1/skills-agent --skill deepsearch_service

Review the files below or copy the command above to add this skill to your agents.

Files (6)

SKILL.md

2.3 KB

---
name: deepsearch-service
description: 基于 LLM 的深度迭代搜索与推理服务。擅长处理复杂问题，通过自动分解查询、多轮迭代检索、信息评估与验证，最终生成全面且结构化的深度分析报告。
---

## 功能
基于 LLM 的迭代式深度搜索服务，能够：
1. 将复杂问题分解为多个子查询
2. 迭代搜索收集多源信息
3. 评估信息充分性，动态调整搜索策略
4. 综合生成结构化分析报告

## 适用场景
- 复杂问题需要多角度分析
- 需要综合多个来源的信息
- 要求生成完整的研究报告

## 调用方式
```python
from services.deepsearch_service.client import DeepSearchClient

client = DeepSearchClient()

# 健康检查
status = client.health_check()

# 深度搜索（默认参数）
result = client.search("Python异步编程的最佳实践有哪些？")

# 自定义参数
result = client.search(
    query="如何设计一个高可用的微服务架构？",
    max_iterations=3,          # 最大迭代次数 (1-5)
    queries_per_iteration=3,   # 每轮查询数 (1-5)
    depth_level="deep"         # 搜索深度: quick/normal/deep
)

# 获取报告和来源
print(result["report"])
for source in result["sources"]:
    print(f"- {source['title']}: {source['url']}")
```

## 返回格式
```json
{
  "query": "Python异步编程的最佳实践有哪些？",
  "report": "# Python异步编程最佳实践\n\n## 1. 核心概念...",
  "sources": [
    {
      "title": "Python官方asyncio文档",
      "url": "https://docs.python.org/...",
      "relevance": 0.95,
      "snippet": "asyncio是Python标准库中的异步I/O框架..."
    }
  ],
  "iterations": [
    {
      "iteration": 1,
      "queries": ["Python asyncio 教程", "async await 用法"],
      "results_count": 6,
      "key_findings": ["asyncio是标准库组件"]
    }
  ],
  "total_iterations": 2,
  "total_sources": 8,
  "elapsed_seconds": 45.32,
  "search_timestamp": "2025-12-28T18:30:00"
}
```

## 与 websearch_service 的区别

| 特性 | WebSearch | DeepSearch |
|------|-----------|------------|
| 搜索模式 | 单轮搜索 | 迭代多轮搜索 |
| 查询策略 | 直接使用输入 | LLM 分解/优化 |
| 输出格式 | 结构化列表 | 综合报告 + 来源 |
| 响应时间 | 10-30s | 30-120s |

Overview

This skill provides an LLM-driven deep iterative search and reasoning service that breaks down complex questions, gathers multi-source evidence, and delivers structured analytical reports. It is designed to adapt search strategy across multiple iterations to improve completeness and relevance. The output includes a comprehensive report, provenance metadata, and iteration logs for transparency.

How this skill works

The service uses an LLM to decompose the input query into focused subqueries, runs multiple retrieval passes, and aggregates results from diverse sources. After each iteration it evaluates information sufficiency and dynamically adjusts queries and depth. Final output is a cohesive, structured report with ranked sources, iteration history, and timing metrics.

When to use it

Investigating complex technical or strategic problems requiring multi-angle analysis
Producing comprehensive research or briefing reports with source attribution
When initial single-pass search yields incomplete or inconsistent information
Validating hypotheses by iteratively refining queries and evidence
Compiling summaries from heterogeneous sources for decision support

Best practices

Start with a clear, scoped question and provide key constraints or goals
Adjust max_iterations and queries_per_iteration for depth vs latency trade-offs
Specify depth_level (quick/normal/deep) to match research needs and time budget
Review iteration logs and source relevance scores to understand how conclusions were formed
Combine the generated report with domain expert review for high-stakes decisions

Example use cases

Designing a high-availability microservices architecture with cross-source justification
Creating a literature-style review on best practices for Python async programming
Investigating regulatory impacts across jurisdictions with sourced evidence
Generating an executive briefing that synthesizes technical, business, and risk factors
Validating competing claims by tracing evidence across iteration history

FAQ

How long does a deep search typically take?

Response time varies by depth_level and iterations; expect roughly 30–120 seconds for deep runs and 10–30 seconds for quick checks.

Can I control how many iterations or queries run?

Yes. You can set max_iterations (1–5) and queries_per_iteration (1–5) to tune thoroughness and latency.