web

Files (1)

• SKILL.md

SKILL.md (first lines)

# Web Skill

## Purpose

Conduct comprehensive web research using multiple search providers with intelligent result synthesis, saturation detection, and HyDE-powered query expansion. For local data, use `/find`.

## Capabilities

### Providers (working)
- **tavily** - Fast, comprehensive search (recommended default)
- **exa** - High-quality technical/engineering results
- **glm** - Web search via GLM-4.5-Air model
- **serpapi** - Google results via serpapi.com (250/month free)
- **serper** - Google results via serper.dev (requires SERPER_API_KEY; backend registered in router_async, not a direct `--mode`)
- **webreader** - URL fetching and content extraction (not a search engine)
- **github** - Code and repository search (requires GITHUB_TOKEN)

### Providers (non-functional)
- **notebooklm** - Document analysis (requires NotebookLM MCP + notebooks)
- **claude** - Claude API documentation (requires ANTHROPIC_API_KEY)

### Modes
- **auto** - Auto-select providers based on query type
- **web** - Web search with multiple providers
- **quick** - Fast results from single provider

### Features
- **HyDE integration** - Query expansion for better results
- **Saturation detection** - Stop when results are semantically saturated
- **URL fetching** - Extract full content from search results
- **Result caching** - Avoid duplicate searches
- **Multi-provider synthesis** - Combine results from multiple sources

## Your Workflow

1. **Analyze query intent** - Detect if technical, general, or code search
2. **Select mode** - Choose auto/web/quick based on query complexity
3. **Choose providers** - Pick appropriate providers for query type
4. **Execute search** - Parallel queries with saturation detection
5. **Synthesize results** - Combine and rank from multiple providers
6. **Fetch URLs** - Extract content when results need verification
7. **Format output** - Present findings with sources and relevance scores

## Provider Selection Guide

| Query Type | Recommended Providers | Alternative |
|------------|----------------------|-------------|
| **General research** | tavily | exa, serpapi |
| **Technical/Code** | exa, github | - |
| **AI/ML topics** | tavily, exa | - |
| **Chinese content** | glm | - |
| **Documentation** | webreader + tavily | - |
| **Fast results** | tavily only | quick mode |

## Usage Examples

### Basic research
```
/web "async best practices in Python"
```

### Technical search with specific provider
```
/web "FastAPI vs Flask performance" --mode exa
```
(Note: Provider names are used as modes, not as separate --providers flag)

### Quick search
```
/web "numpy array operations" --mode quick
```

### Research with URL fetching
```
/web "microservices architecture" --fetch-urls 10
```

### Code-specific search
```
/web "pytest fixtures tutorial" --mode github
```
(Note: Provider names are used as modes)

## GitHub Search Routing

### Intelligent Code vs. Repository Detection

When using `--mode github`, the system intelligently routes between **code search** and **repository search** based on query patterns:

**Routes to CODE SEARCH when**:
- **Code keywords detected**: `function`, `class`, `def`, `import`, `async def`, `await`, `const`, `let`, `var`, `interface`, `component`, `hook`, `lambda`, `decorator`, etc.
- **File extensions present**: `.py`, `.js`, `.ts`, `.java`, `.go`, `.rs`, `.cpp`, `.dockerfile`, `.json`, `.yaml`, etc.
- **Language specifiers used**: `language:python`, `language:javascript`, etc. (GitHub API syntax)

**Routes to REPOSITORY SEARCH when**:
- **No code patterns detected**: General queries like "pytest testing libraries"
- **Repository-focused terms**: "framework", "library", "package", "tool", "tutorial", "guide", "comparison"
- **Ambiguous queries**: Defaults to repository search (safe fallback)

### Known Limitations

**False Positives** (documented in pre-mortem analysis):
- ❌ `"python async framework"` → Routes to code search (user wants repos)
- ❌ `"import tutorial"` → Routes to code search (user wants tutorials)

**These occur because**: Current implementation uses keyword presence without phrase-level context analysis.

**Workaround**: Use web providers for repo discovery when you encounter false positives:
```
/web "python async framework" --mode tavily
```

### Rate Limit Handling

**Automatic fallback**: If code search is rate-limited by GitHub API, the system automatically falls back to repository search with a warning:
```
Warning: "Code search rate limited - showing repositories instead"
```

**Monitoring**: Rate limit errors are logged for monitoring and debugging.

From EndUser123/search-research, MIT.

More from EndUser123/search-research