Skip to main content

​
MCP Integration

The Aris RAG pipeline is fully compliant with the Model Context Protocol (MCP). This allows any MCP-enabled agent (Claude Desktop, internal agents) to utilize our vector stores and knowledge graphs as standard tools.

​
Tool Definitions

We expose two primary tools via the MCP server: retrieve_context and semantic_search. 
{
  "name": "retrieve_context",
  "description": "Retrieves high-precision context for answering a user query using Hybrid Search and RRF.",
  "inputSchema": {
    "type": "object",
    "properties": {
      "query": {
        "type": "string",
        "description": "The user's natural language query."
      },
      "filters": {
        "type": "object",
        "description": "Metadata filters (e.g., source, author, date range).",
        "properties": {
          "source": { "type": "string" },
          "tags": { "type": "array", "items": { "type": "string" } }
        }
      },
      "top_k": {
        "type": "integer",
        "description": "Number of chunks to retrieve (default: 10).",
        "minimum": 1,
        "maximum": 50
      }
    },
    "required": ["query"]
  }
}

{
  "name": "semantic_search",
  "description": "Performs a raw vector similarity search without RRF or hybrid fusion. Use for exploration.",
  "inputSchema": {
    "type": "object",
    "properties": {
      "vector_query": {
        "type": "string",
        "description": "The text to embed and search."
      },
      "threshold": {
        "type": "number",
        "description": "Minimum similarity score (0.0 to 1.0).",
        "default": 0.7
      }
    },
    "required": ["vector_query"]
  }
}

​
Server Implementation

Our MCP server is built using the mcp-python-sdk. It connects to the configured vector database (Qdrant/Pinecone) and manages the FastAPI lifecycle.

MCP Server Lifecycle

  1. Initialize: Connect to Vector DB and embedding model.
  2. Expose Tools: Register retrieve_context and semantic_search capabilities.
  3. Handle Requests: Validate JSON-RPC 2.0 requests.
  4. Execute & Return: Run the RAG pipeline and return CallToolResult.
# Server-side implementation snippet
@mcp.tool()
async def retrieve_context(query: str, filters: dict = None, top_k: int = 10) -> str:
    """
    Executes the hybrid retrieval pipeline.
    """
    ctx.request_id = generate_uuid()
    
    # 1. Expand Query
    expanded_query = await query_rewriter.rewrite(query)
    
    # 2. Retrieve
    results = await retriever.hybrid_search(
        query=expanded_query,
        filters=filters,
        k=top_k
    )
    
    # 3. Format as Context String
    return format_results(results)
Ensure your MCP server has access to the same environment variables (OPENAI_API_KEY, VECTOR_DB_URL) as the main application.
Last modified on February 19, 2026