Query Assistant MCP Server

A Model Context Protocol (MCP) server that helps generate queries using semantic search over training examples. This server provides few-shot learning capabilities by finding similar questions and their corresponding queries to guide query generation for various query languages (Cypher, SPARQL, SQL, etc.).

Features

• Semantic Search: Uses OpenAI embeddings to find similar questions in your training dataset
• Few-Shot Learning: Returns relevant examples to help generate accurate queries
• Training Data Management: Add, list, and manage question-query pairs with duplicate detection
• Vector Storage: Efficient similarity search using HNSW (Hierarchical Navigable Small World) algorithm
• Metadata Support: Organize examples by domain, complexity, and tags
• Multi-Language Support: Works with various query languages (Cypher, SPARQL, SQL, etc.)

Installation

1. Prerequisites: Node.js 18+ and npm

2. Build the server:

   cd /Users/alkhalili/Documents/Cline/MCP/mcp-query-assistant
   npm install
   npm run build

3. Configure OpenAI API Key:

   • Get your API key from OpenAI Platform
   • Update the MCP settings file at: /Users/alkhalili/Library/Application Support/Code/User/globalStorage/saoudrizwan.claude-dev/settings/cline_mcp_settings.json
   • Replace your_openai_api_key_here with your actual API key

4. Configure Data Directory (Optional):

   • By default, data is stored in the data/ folder relative to the project
   • To use a custom data directory, you can:
     • Set the DATA_DIR environment variable, or
     • Use the --data-dir command line argument

5. Server Configuration: The server is already configured in your MCP settings as:

   "query-assistant": {
     "command": "npx",
     "args": ["mcp-query-assistant"],
     "env": {
       "OPENAI_API_KEY": "your_openai_api_key_here",
       "DATA_DIR": "/optional/custom/data/path"
     },
     "disabled": false,
     "autoApprove": []
   }

   Alternative with command line arguments:

   "query-assistant": {
     "command": "npx",
     "args": ["mcp-query-assistant", "--openai-key", "your_openai_api_key_here", "--data-dir", "/custom/data/path"],
     "disabled": false,
     "autoApprove": []
   }

Available Tools

1. find_similar_queries

Find similar query examples based on a natural language question.

Parameters:

• question (required): The natural language question to find similar examples for
• limit (optional): Maximum number of similar examples to return (default: 3, max: 10)
• threshold (optional): Minimum similarity threshold 0-1 (default: 0.7)

Example Usage:

Use the find_similar_queries tool to find examples for: "Give me the list of CDEs in the lineage"

2. add_training_example

Add a new question-query pair to the training dataset.

Parameters:

• question (required): The natural language question
• query (required): The corresponding query (Cypher, SPARQL, SQL, etc.)
• metadata (optional): Additional metadata (domain, complexity, tags)

Example Usage:

Add a training example:
Question: "Find users who bought expensive products"
Query: "MATCH (u:User)-[:PURCHASED]->(p:Product) WHERE p.price > 1000 RETURN u"
Metadata: {"domain": "user_analytics", "complexity": "medium"}

3. list_training_examples

List all training examples in the dataset.

Parameters:

• limit (optional): Maximum number of examples to return (default: 10, max: 100)
• domain (optional): Filter by domain

4. find_duplicates

Find duplicate training examples based on question and query.

Parameters: None

Example Usage:

Use find_duplicates to identify duplicate training examples in your dataset.

5. remove_duplicates

Remove duplicate training examples, keeping only the first occurrence of each unique question-query pair.

Parameters:

• confirm (optional): Set to true to confirm removal of duplicates (default: false)

Example Usage:

Use remove_duplicates with confirm=true to clean up duplicate examples.

Default Training Examples

The server comes with 1 default example covering data lineage patterns:

1. Data Lineage: "Give me the list of CDEs in the lineage"
   • Query: MATCH (cde:CDE) RETURN cde.name, cde.description, cde.layer, cde.fqn ORDER BY cde.name
   • Domain: Data Lineage
   • Complexity: simple

Usage Workflow

1. Ask for similar queries: When you need to write a query, use find_similar_queries with your natural language question
2. Get few-shot examples: The server returns similar questions and their queries with similarity scores
3. Generate your query: Use the examples as guidance to write your specific query
4. Add new examples: Use add_training_example to expand your training dataset with new patterns
5. Manage duplicates: Use find_duplicates and remove_duplicates to keep your dataset clean

Data Storage

• Training Data: Stored in {DATA_DIR}/training_data.json (default: data/training_data.json)
• Vector Index: Stored in {DATA_DIR}/vector_index.bin (default: data/vector_index.bin)
• Embeddings: Generated using OpenAI's text-embedding-3-small model (1536 dimensions)
• Data Directory: Configurable via DATA_DIR environment variable or --data-dir command line argument

Example Interaction

User: "Show me all data elements with their descriptions"

Agent: Let me find similar examples for you.
[Uses find_similar_queries tool]

Server Response:
Found 1 similar examples for: "Show me all data elements with their descriptions"

Example 1 (similarity: 0.823):
Question: Give me the list of CDEs in the lineage
Query: MATCH (cde:CDE) RETURN cde.name, cde.description, cde.layer, cde.fqn ORDER BY cde.name
Domain: Data Lineage
Complexity: simple

Agent: Based on this example, here's a query for showing data elements with descriptions:
MATCH (cde:CDE) 
RETURN cde.name, cde.description 
ORDER BY cde.name

Troubleshooting

1. "OpenAI API key not configured": Make sure you've set your API key in the MCP settings
2. "No similar examples found": Try lowering the similarity threshold or adding more training data
3. Server not connecting: Check that the build path is correct and the server compiled successfully

Contributing

To add more training examples or improve the server:

1. Use the add_training_example tool to add new question-query pairs
2. Organize examples with appropriate metadata (domain, complexity, tags)
3. Test similarity search with various question phrasings
4. Consider adding domain-specific examples for your use case

Technical Details

• Vector Database: HNSW (Hierarchical Navigable Small World) for efficient similarity search
• Embedding Model: OpenAI text-embedding-3-small (1536 dimensions)
• Similarity Metric: Cosine similarity
• Storage Format: JSON for training data, binary for vector index
• Max Capacity: 10,000 training examples (configurable)