grok-search

name	grok-search
description	Search the web for real-time information using dual engines: Grok AI (deep analysis + live web) and Tavily (structured results + web scraping). Use when needing current news, prices, latest releases, or any info beyond training data cutoff. Chinese triggers: 搜索, 搜一下, 查一下, 查查, 看看有没有, 最新, 新闻, grok搜索, 网上搜, 帮我查, 了解一下, 什么价格, 上线了吗, 发布了吗.

Prerequisites

Set these environment variables before use:

• GROK_API_BASE — Your Grok-compatible API endpoint (e.g., http://localhost:8100/v1)
• GROK_API_KEY — Your API key for the Grok endpoint
• TAVILY_API_KEY — Your Tavily API key (get one free at tavily.com)

Instructions

Search for: $ARGUMENTS

---

Step 1: Do I Need to Search?

Yes — search when:

• User explicitly asks for external info
• Real-time data needed (latest versions, recent events, current prices)
• Need to verify knowledge accuracy
• Specific URLs, projects, or products' current status

No — skip when:

• Pure code writing/debugging with sufficient context
• User explicitly says not to search

---

Step 2: Choose Engine

Scenario	Engine	Why
Simple fact lookup	Tavily	Fast, structured, scored results
Complex / controversial topic	Dual (both)	Cross-verify, reduce hallucination
Need AI deep analysis	Grok	Built-in web search + reasoning
Latest news	Tavily (topic=news)	News mode optimized
Scrape specific URL	Tavily Extract	Full page content as Markdown
High-stakes decision	Dual (both)	≥2 sources required

---

Step 3: Execute

Grok Search (AI-driven, deep analysis)

curl -s ${GROK_API_BASE}/chat/completions \
  -H "Content-Type: application/json" \
  -H "Authorization: Bearer ${GROK_API_KEY}" \
  -d '{
    "model": "grok-3",
    "messages": [
      {"role": "system", "content": "You are a search assistant. Search the web and provide accurate, well-sourced answers."},
      {"role": "user", "content": "SEARCH_QUERY"}
    ],
    "stream": false
  }' | jq -r '.choices[0].message.content'

Tavily Search (structured results)

curl -s "https://api.tavily.com/search" \
  -H "Content-Type: application/json" \
  -d '{
    "api_key": "'${TAVILY_API_KEY}'",
    "query": "SEARCH_QUERY",
    "search_depth": "basic",
    "max_results": 5,
    "include_answer": true
  }' | jq '{answer: .answer, results: [.results[] | {title, url, content, score}]}'

Options:

• search_depth: basic (fast) or advanced (thorough)
• topic: general (default), news, finance
• max_results: 1-20
• time_range: day, week, month, year

Tavily Extract (scrape a URL)

curl -s "https://api.tavily.com/extract" \
  -H "Content-Type: application/json" \
  -d '{
    "api_key": "'${TAVILY_API_KEY}'",
    "urls": ["https://example.com/page"],
    "extract_depth": "basic"
  }' | jq '.results[] | {url, raw_content}'

Dual Search (cross-verify)

Run both in parallel for important queries, then compare results.

---

Evidence Standards

• All factual conclusions need ≥2 independent sources for cross-verification
• If relying on a single source, explicitly state this limitation
• Prefer: official docs, Wikipedia, academic databases, authoritative media
• When sources conflict: present both sides, assess credibility and recency
• Confidence levels: High (multiple sources agree) · Medium (few sources) · Low (single source)

Search Complexity

Level	Description	Strategy
L1 - Simple	Single fact, one query	Single engine, direct answer
L2 - Medium	Multiple aspects	Primary engine + follow-up
L3 - Complex	Multi-faceted	Dual search + structured planning