skill-web-search

name	skill-web-search
description	Web search + deduplication + summary extraction for investment report evidence gathering. Public deployment uses Volcengine/Baidu; intranet uses Nova MCP web_search.

Web Search + Evidence Extraction

Purpose

Search for supplementary evidence beyond the RAG database.

Tool Selection

部署环境	工具	说明
公网	ivr_web_search	火山引擎 / 百度搜索
内网	Nova MCP web_search	MCP Server 端口 8092，自动发现

Agent 启动时会自动检测可用工具。ivr_web_search 不可用（内网 IVR_SEARCH_ENABLED=false）时，使用 Nova MCP web_search。无需手动选择。

Process

1. Search

• 公网：执行 ivr_web_search，携带 keywords + company name + industry
• 内网：调用 Nova MCP web_search 工具，入参 query（搜索关键词）+ limit（默认 5）
• 优先引用权威信源：政府网站、行业协会、研究机构、官方公告

2. Fetch & Extract

• Call web_fetch on each search result URL
• Extract: title, publication date, relevant paragraph, data points
• Mark source credibility: High (government/official), Medium (industry report), Low (media/blog)

3. Deduplicate

• Remove results with identical URLs
• Merge results covering the same data point (keep highest-credibility source)

4. Output Format

For each result:

• title: page/article title
• url: source URL
• snippet: relevant extracted text (max 300 chars)
• credibility: high | medium | low
• date: publication date if available

Anti-Patterns

• Do NOT cite sources without verifying the URL is accessible
• Do NOT use paywalled content (no way to verify)
• Do NOT treat web search as primary source when RAG has PDF evidence on the same topic