Qdrant MCP Server

A Model Context Protocol (MCP) server that provides semantic search capabilities using a local Qdrant vector database and OpenAI embeddings.

Features

• Semantic Search: Natural language search across your document collections

• Metadata Filtering: Filter search results by metadata fields using Qdrant's powerful filter syntax

• Local Vector Database: Runs Qdrant locally via Docker for complete data privacy

• Automatic Embeddings: Uses OpenAI's embedding models to convert text to vectors

• Rate Limiting: Intelligent request throttling with exponential backoff to prevent API rate limit errors

• MCP Integration: Works seamlessly with Claude Code and other MCP clients

• Collection Management: Create, list, and delete vector collections

• Document Operations: Add, search, and delete documents with metadata support

Prerequisites

• Node.js 18+

• Docker and Docker Compose

• OpenAI API key

Installation

1. Clone the repository:

2. Install dependencies:

3. Set up environment variables:

Edit .env and add your OpenAI API key:

4. Start Qdrant:

5. Build the project:

Usage

Running the Server

For development:

For production:

Claude Code Configuration (Linux)

Add this to your Claude Code configuration file at ~/.claude/claude_code_config.json:

Replace YOUR_USERNAME and the path with your actual username and installation path.

Restart Claude Code after making this change.

Available Tools

create_collection

Create a new vector collection.

Parameters:

• name (string, required): Collection name

• distance (string, optional): Distance metric - "Cosine", "Euclid", or "Dot" (default: "Cosine")

Example:

add_documents

Add documents to a collection with automatic embedding generation.

Parameters:

• collection (string, required): Collection name

• documents (array, required): Array of documents with:

  • id (string/number, required): Unique identifier (string IDs are automatically normalized to UUID format)

  • text (string, required): Document text content

  • metadata (object, optional): Additional metadata

Note: String IDs are automatically normalized to UUID format for Qdrant compatibility. The normalization is deterministic, so the same string ID will always produce the same UUID.

Example:

semantic_search

Search for documents using natural language.

Parameters:

• collection (string, required): Collection to search

• query (string, required): Search query

• limit (number, optional): Max results (default: 5)

• filter (object, optional): Metadata filter in Qdrant format

Filter Format:

The filter parameter accepts Qdrant's native filter format for powerful metadata-based filtering:

You can also use more complex filters:

• Multiple conditions (AND): Use must with multiple conditions

• Any condition (OR): Use should with multiple conditions

• Negation (NOT): Use must_not with conditions

Example with multiple conditions:

Examples:

Basic search:

With single filter:

With multiple filters (AND):

list_collections

List all available collections.

get_collection_info

Get detailed information about a collection.

Parameters:

• name (string, required): Collection name

delete_collection

Delete a collection and all its documents.

Parameters:

• name (string, required): Collection name

delete_documents

Delete specific documents from a collection.

Parameters:

• collection (string, required): Collection name

• ids (array, required): Array of document IDs to delete (string IDs are automatically normalized to UUID format)

Note: String IDs will be normalized to the same UUID format used when adding documents.

Available Resources

• qdrant://collections - List all collections

• qdrant://collection/{name} - Collection details and statistics

Project Structure

Example Workflow

1. Create a collection:

   Create a collection called "knowledge-base"

2. Add documents:

   Add these documents to knowledge-base: - id: "doc1", text: "MCP is a protocol for AI model context", metadata: {"type": "definition", "category": "protocol"} - id: "doc2", text: "Vector databases store embeddings for semantic search", metadata: {"type": "definition", "category": "database"} - id: "doc3", text: "Qdrant provides high-performance vector similarity search", metadata: {"type": "product", "category": "database"}

3. Search without filters:

   Search knowledge-base for "how does semantic search work"

4. Search with filters:

   Search knowledge-base for "vector database" with filter {"must": [{"key": "category", "match": {"value": "database"}}]}

5. Get collection information:

   Get info about "knowledge-base" collection

6. View all collections:

   What collections do I have?

Configuration Options

Environment Variables

• OPENAI_API_KEY (required): Your OpenAI API key

• QDRANT_URL (optional): Qdrant server URL (default: http://localhost:6333)

• OPENAI_EMBEDDING_MODEL (optional): Embedding model (default: text-embedding-3-small)

• OPENAI_EMBEDDING_DIMENSIONS (optional): Custom embedding dimensions

• OPENAI_MAX_REQUESTS_PER_MINUTE (optional): Maximum OpenAI API requests per minute (default: 3500)

• OPENAI_RETRY_ATTEMPTS (optional): Number of retry attempts for rate limit errors (default: 3)

• OPENAI_RETRY_DELAY (optional): Initial retry delay in milliseconds with exponential backoff (default: 1000)

Embedding Models

Available OpenAI models:

• text-embedding-3-small (1536 dims, faster, cheaper)

• text-embedding-3-large (3072 dims, higher quality)

Advanced Features

Rate Limiting and Error Handling

The server implements robust rate limiting to handle OpenAI API limits gracefully:

Features:

• Request Throttling: Queues requests to stay within OpenAI's rate limits (configurable, default: 3500 requests/minute)

• Exponential Backoff: Automatically retries failed requests with increasing delays (1s, 2s, 4s, 8s...)

• Retry-After Header Support: Respects OpenAI's retry guidance with validation fallback for invalid headers

• Typed Error Handling: Uses OpenAIError interface for type-safe error detection and handling

• Smart Error Detection: Identifies rate limit errors (429 status) vs other failures

• User Feedback: Clear console messages during retry attempts with estimated wait times

Configuration:

Benefits:

• Prevents failed operations during high-volume usage

• Automatic recovery from temporary API issues

• Optimized for batch document processing

• Works seamlessly with both single and batch embedding operations

Metadata Filtering

The server supports Qdrant's powerful filtering capabilities for refined search results. Filters can be applied to any metadata field stored with your documents.

Supported filter types:

• Match filters: Exact value matching for strings, numbers, and booleans

• Logical operators: must (AND), should (OR), must_not (NOT)

• Range filters: Greater than, less than, between (for numeric values)

• Nested filters: Complex boolean expressions

See the semantic_search tool documentation for filter syntax examples.

Troubleshooting

Qdrant connection errors

Make sure Qdrant is running:

If not running:

Collection doesn't exist

Create the collection first before adding documents:

OpenAI API errors

• Verify your API key is correct in .env

• Check your OpenAI account has available credits

• Ensure you have access to the embedding models

Rate limit errors

The server automatically handles rate limits, but if you see persistent rate limit errors:

• Reduce OPENAI_MAX_REQUESTS_PER_MINUTE to match your OpenAI tier (free: 500, paid: 3500+)

• Increase OPENAI_RETRY_ATTEMPTS for more resilient retries

• Check your OpenAI dashboard for current usage and limits

Filter errors

If you encounter "Bad Request" errors with filters:

• Ensure you're using Qdrant's native filter format

• Check that field names match your metadata exactly

• Verify the filter structure has proper nesting

Development

Development Mode

Run in development mode with auto-reload:

Build

Build for production:

Type Checking

Run TypeScript type checking without emitting files:

Continuous Integration

The project uses GitHub Actions for CI/CD:

• Build: Compiles TypeScript to JavaScript

• Type Check: Validates TypeScript types with strict mode

• Test: Runs all 140 unit and functional tests (129 unit + 11 functional)

• Multi-version: Tests on Node.js 18, 20, and 22

The CI workflow runs on every push and pull request to the main branch.

Testing

The project includes comprehensive unit and integration tests using Vitest.

Run Tests

Test Coverage

The test suite includes 140 tests (129 unit + 11 functional) covering:

Unit Tests:

• QdrantManager (21 tests): Collection management, point operations, and search functionality

• OpenAIEmbeddings (25 tests): Embedding generation, batch processing, rate limiting with exponential backoff, Retry-After header validation, and typed error handling

• MCP Server (19 tests): Tool schemas, resource URI patterns, and MCP protocol compliance

Functional Tests:

• Live API Integration (11 tests): Real OpenAI embeddings, production MCP server validation, rate limiting behavior, and end-to-end workflows with 30+ real documents

Coverage Highlights:

• ✅ 100% function coverage across all modules

• ✅ Comprehensive rate limiting tests with timing validation

• ✅ Typed error handling with OpenAIError interface

• ✅ Invalid Retry-After header fallback to exponential backoff

• ✅ Real-world validation with live OpenAI API

• ✅ Both source and compiled code tested

See docs/test_report.md for detailed test results and coverage analysis.

Writing Tests

Tests are located next to the files they test with a .test.ts extension:

• src/qdrant/client.test.ts - Qdrant client wrapper tests

• src/embeddings/openai.test.ts - OpenAI embeddings provider tests

• src/index.test.ts - MCP server integration tests

Run tests before committing:

License

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

Contributing

Contributions are welcome! Please feel free to submit a Pull Request.

Before Submitting a PR

1. Run tests: Ensure all tests pass

   npm test -- --run

2. Type check: Verify no TypeScript errors

   npm run type-check

3. Build: Confirm the project builds successfully

   npm run build

All pull requests will automatically run through CI checks that validate:

• TypeScript compilation

• Type checking

• Test suite (114 tests)

• Compatibility with Node.js 18, 20, and 22

Note for Repository Owners

After forking or cloning, update the CI badge URL in README.md:

Replace YOUR_USERNAME with your GitHub username or organization name.