QDrant Loader MCP Server

A Model Context Protocol (MCP) server that brings advanced RAG search to AI development tools. Part of the QDrant Loader monorepo ecosystem.

🎯 What It Does

• Provides intelligent search through semantic, hierarchy-aware, and attachment-focused tools
• Integrates seamlessly with Cursor, Windsurf, Claude Desktop, and other MCP-compatible tools
• Understands context including document hierarchies, file relationships, and metadata
• Streams responses for fast, real-time search results
• Preserves relationships between documents, attachments, and parent content

🔌 Supported AI Tools

For per-tool JSON configuration, see MCP setup and integration.

🔍 Search Tools

Core search tools

Search Intelligence Features

• Hierarchy Understanding: Recognizes parent/child page relationships in Confluence
• Attachment Awareness: Connects files to their parent documents and context
• Metadata Enrichment: Includes authors, dates, file sizes, and source information
• Visual Indicators: Rich formatting with icons and context clues
• Relationship Mapping: Shows connections between related content

Additional MCP Tools

Beyond the three core search tools, the server also provides cross-document and expansion tools:

analyze_relationships, find_similar_documents, detect_document_conflicts, find_complementary_content, cluster_documents, expand_document, expand_cluster, expand_chunk_context

For full parameter references and usage examples, see MCP search capabilities.

📦 Installation

pip install qdrant-loader-mcp-server

For the full ingestion + MCP pipeline:

pip install qdrant-loader qdrant-loader-mcp-server

🚀 Quick Start

1. Set environment variables

export QDRANT_URL="http://localhost:6333"
export QDRANT_API_KEY="your_api_key"        # Required for QDrant Cloud
export LLM_API_KEY="your_openai_key"

# Optional configuration
export QDRANT_COLLECTION_NAME="documents"  # Default collection name
export MCP_LOG_LEVEL="INFO"                # Logging level
export MCP_LOG_FILE="/path/to/mcp.log"     # Log file path
export MCP_DISABLE_CONSOLE_LOGGING="true"  # Recommended for Cursor

2. Start the server

mcp-qdrant-loader

# With debug logging
mcp-qdrant-loader --log-level DEBUG

# Help
mcp-qdrant-loader --help

3. Test the connection

# Test with a simple search
echo '{"jsonrpc":"2.0","id":1,"method":"tools/call","params":{"name":"search","arguments":{"query":"test","limit":1}}}' | mcp-qdrant-loader

🔧 Configuration

Environment variables

For Cursor: Always set MCP_DISABLE_CONSOLE_LOGGING=true to prevent JSON-RPC interference. Use MCP_LOG_FILE to capture logs instead.

HTTP transport and workers

For production deployments with multiple worker processes:

# Start with HTTP transport (single worker, good for local development)
mcp-qdrant-loader --transport http --port 8080

# Start with multiple workers for production
mcp-qdrant-loader --transport http --port 8080 --workers 4

Each worker is a separate OS process with its own event loop, Qdrant connection pool, and search engine. This eliminates GIL contention for CPU-bound work (SpaCy, BM25, reranking).

Tuning Concurrency

SEARCH_MAX_CONCURRENT limits the number of simultaneous Qdrant queries per worker. With multiple workers, the total concurrent load on Qdrant is workers × SEARCH_MAX_CONCURRENT.

If you see 408 Request Timeout from Qdrant, lower SEARCH_MAX_CONCURRENT to match your Qdrant instance's capacity:

export SEARCH_MAX_CONCURRENT=2
mcp-qdrant-loader --transport http --workers 4

🎯 Usage Examples

Ask your AI assistant in natural language:

• "Find documentation about authentication in our API"
• "Show me examples of error handling patterns in our codebase"
• "What are the deployment requirements for this service?"
• "Find all PDF attachments related to database schema"
• "Show me the hierarchy of pages under the Architecture section"

Advanced Search Queries

Semantic Search

Find information about rate limiting implementation

Hierarchy Search

Show me all child pages under the API Documentation section

Attachment Search

Find all Excel files uploaded by john.doe in the last month

🏗️ Architecture

MCP Protocol Implementation

• Full MCP 2024-11-05 compliance with proper JSON-RPC communication
• Tool registration for search, hierarchy_search, and attachment_search
• Streaming responses for large result sets
• Error handling with proper MCP error codes
• Resource management for efficient memory usage

Search Engine Components

• Embedding Service — Generates query embeddings using the configured LLM provider
• Vector Search — Performs semantic similarity search in QDrant
• Metadata Processor — Enriches results with hierarchy and attachment information
• Result Formatter — Creates rich, contextual response formatting
• Caching Layer — Optimizes performance for repeated queries

Data Flow

AI Tool → MCP Server → QDrant Search → Result Processing → Formatted Response
    ↓         ↓            ↓              ↓                ↓
Cursor    JSON-RPC    Vector Query   Metadata         Rich Context
Windsurf  Protocol    Embedding      Enrichment       Visual Indicators
Claude    Tool Call   Similarity     Hierarchy        Relationship Info
Other     Streaming   Ranking        Attachments      Source Attribution

For system-level architecture, see Architecture guide.

📚 Documentation

• Getting Started - Quick start and core concepts
• Monorepo overview - Repository layout, package responsibilities, and navigation entry points.
• Quick start - Fast setup path from install to first successful ingestion.
• User Guides - Detailed usage instructions
• Developer Docs - Architecture and API reference
• MCP server guide - End-to-end overview of MCP usage with QDrant Loader.
• Basic Configuration - Getting started with configuration

🆘 Support

• Issues - Bug reports and feature requests
• Discussions - Community Q&A

🤝 Contributing

See CONTRIBUTING - Contribution guidelines, development standards, and pull request process.

📄 License

This project is licensed under the Apache License 2.0 - see the LICENSE file for details.

Ready to get started? Check out our Quick Start Guide or browse the complete documentation.