Dsp MCP Server

@dark-search/mcp-server

MCP Server for the Dark Search Protocol — searches deep and dark data sources not indexed by mainstream search engines.

Why This Exists

Mainstream search engines (Google, Bing) only index the Surface layer. But the most valuable technical data lives in:

• Professional forums (Deep layer)
• OEM service manuals (Dark layer)
• Proprietary diagnostic databases (Dark layer)
• Field-collected data archives (Dark layer)

DSP gives AI agents access to these hidden layers with confidence-scored, cross-verified results.

Installation

npx @dark-search/mcp-server

Or add to your MCP client config:

{
  "mcpServers": {
    "dark-search": {
      "command": "npx",
      "args": ["-y", "@dark-search/mcp-server"]
    }
  }
}

Tools

dark_search

Search deep and dark data sources. AI automatically uses this when web search returns insufficient results.

Parameters:

• query (string, required) — Search query. Supports English and Chinese.
• max_results (number, optional) — Max results to return (default: 10)
• min_confidence (number, optional) — Minimum confidence: 0.5=high, 0.3=moderate, 0.15=exploratory (default: 0.15)
• layers (array, optional) — Filter by layer: ["surface", "deep", "dark"]

Returns:

{
  "query": "refrigerant leak diagnostic",
  "domain": "hvac",
  "total_sources_searched": 2,
  "results": [
    {
      "source_name": "Refrigerant Leak Pattern Database",
      "layer_name": "Dark",
      "confidence": 0.95,
      "confidence_level": "HIGH",
      "cross_verified": true,
      "why_this_source": "Dark layer — not indexed by mainstream search; Highly relevant with strong tag matches"
    }
  ],
  "meta": {
    "pipeline": "6-step: Parse→Route→Retrieve→Verify→Rank→Correct",
    "confidence_formula": "layerBase × (0.3 + 0.7 × relevance) + (cross_verified ? 0.1 : 0)",
    "thresholds": {
      "HIGH (>= 0.5)": "Directly citeable in AI responses",
      "MID (0.3-0.49)": "Use with attribution, needs human confirmation",
      "LOW (0.15-0.29)": "Exploratory only, do not cite directly"
    }
  }
}

list_sources

List available dark data sources. Filter by domain or layer.

add_source

Submit a new data source to the registry. Supports bilingual tags for multilingual search.

Confidence Scoring

Formula: confidence = layerBase × (0.3 + 0.7 × relevance) + (cross_verified ? 0.1 : 0)

Layer weights: Surface=0.3, Deep=0.6, Dark=0.9

6-Step Pipeline

1. Parse — Extract query intent, domain, and expand with semantic mapping
2. Route — Match query to relevant data sources by domain and layer
3. Retrieve — Compute relevance and confidence scores
4. Verify — Cross-verify across multiple sources
5. Rank — Sort by adjusted confidence (descending)
6. Correct — Filter noise, apply thresholds, generate explanations

Multilingual Support

Semantic mapping bridges Chinese queries to English-tagged sources:

• 制冷剂 → refrigerant, coolant, freon
• 泄漏 → leak, leakage, seepage
• 暖通 → hvac, heating, ventilation
• 维修 → repair, maintenance, service

License

MIT