local-rag-search（本地 RAG 搜索）

Local RAG Search Skill

This skill enables you to effectively use the mcp-local-rag MCP server for intelligent web searches with semantic ranking. The server performs RAG-like similarity scoring to prioritize the most relevant results without requiring any external APIs.

Available Tools

1. rag_search_ddgs - DuckDuckGo Search

Use this for privacy-focused, general web searches.

When to use:

• User prefers privacy-focused searches
• General information lookup
• Default choice for most queries

Parameters:

• query: Natural language search query
• num_results: Initial results to fetch (default: 10)
• top_k: Most relevant results to return (default: 5)
• include_urls: Include source URLs (default: true)

2. rag_search_google - Google Search

Use this for comprehensive, technical, or detailed searches.

When to use:

• Technical or scientific queries
• Need comprehensive coverage
• Searching for specific documentation

3. deep_research - Multi-Engine Deep Research

Use this for comprehensive research across multiple search engines.

When to use:

• Researching complex topics requiring broad coverage
• Need diverse perspectives from multiple sources
• Gathering comprehensive information on a subject

Available backends:

• duckduckgo: Privacy-focused general search
• google: Comprehensive technical results
• bing: Microsoft's search engine
• brave: Privacy-first search
• wikipedia: Encyclopedia/factual content
• yahoo, yandex, mojeek, grokipedia: Alternative engines

Default: ["duckduckgo", "google"]

4. deep_research_google - Google-Only Deep Research

Shortcut for deep research using only Google.

5. deep_research_ddgs - DuckDuckGo-Only Deep Research

Shortcut for deep research using only DuckDuckGo.

Best Practices

Query Formulation

1. Use natural language: Write queries as questions or descriptive phrases

- Good: "latest developments in quantum computing" - Good: "how to implement binary search in Python" - Avoid: Single keywords like "quantum" or "Python"

1. Be specific: Include context and details

- Good: "React hooks best practices for 2024" - Better: "React useEffect cleanup function best practices"

Tool Selection Strategy

1. Single Topic, Quick Answer → Use rag_search_ddgs or rag_search_google rag_search_ddgs(query="What is the capital of France?", top_k=3)
2. Technical/Scientific Query → Use rag_search_google rag_search_google(query="Docker multi-stage build optimization techniques", num_results=15, top_k=7)
3. Comprehensive Research → Use deep_research with multiple search terms deep_research(search_terms=["machine learning fundamentals", "neural networks architecture", "deep learning best practices 2024"], backends=["google", "duckduckgo"], top_k_per_term=5)
4. Factual/Encyclopedia Content → Use deep_research with Wikipedia deep_research(search_terms=["World War II timeline", "WWII key battles"], backends=["wikipedia"], num_results_per_term=5)

Parameter Tuning

For quick answers:

• num_results=5-10, top_k=3-5

For comprehensive research:

• num_results=15-20, top_k=7-10

For deep research:

• num_results_per_term=10-15, top_k_per_term=3-5
• Use 2-5 related search terms
• Use 1-3 backends (more = more comprehensive but slower)

Workflow Examples

Example 1: Current Events

Task: "What happened at the UN climate summit last week?"

1. Use rag_search_google for recent news coverage
2. Set top_k=7 for comprehensive view
3. Present findings with source URLs

Example 2: Technical Deep Dive

Task: "How do I optimize PostgreSQL queries?"

1. Use deep_research with multiple specific terms:
   - "PostgreSQL query optimization techniques"
   - "PostgreSQL index best practices"
   - "PostgreSQL EXPLAIN ANALYZE tutorial"
2. Use backends=["google", "stackoverflow"] if available
3. Synthesize findings into actionable guide

Example 3: Multi-Perspective Research

Task: "Research the impact of remote work on productivity"

1. Use deep_research with diverse search terms:
   - "remote work productivity statistics 2024"
   - "hybrid work model effectiveness studies"
   - "work from home challenges research"
2. Use backends=["google", "duckduckgo"] for broad coverage
3. Synthesize different perspectives and studies

Guidelines

1. Always cite sources: When include_urls=True, reference the source URLs in your response
2. Verify recency: Check if the content appears current and relevant
3. Cross-reference: For important facts, use multiple search terms or engines
4. Respect privacy: Use DuckDuckGo for general queries unless specific needs require Google
5. Batch related queries: When researching a topic, create multiple related search terms for deep_research
6. Semantic relevance: Trust the RAG scoring - top results are semantically closest to the query
7. Explain your choice: Briefly mention which tool you're using and why

Error Handling

If a search returns insufficient results:

1. Try rephrasing the query with different keywords
2. Switch to a different backend
3. Increase num_results parameter
4. Use deep_research with multiple related search terms

Privacy Considerations

• DuckDuckGo: Privacy-focused, doesn't track users
• Google: Most comprehensive but tracks searches
• Recommend DuckDuckGo as default unless user specifically needs Google's coverage

Performance Notes

• First search may be slower (model loading)
• Subsequent searches are faster (cached models)
• More backends = more comprehensive but slower
• Adjust num_results and top_k based on use case