Components Service versus MCP Service
15/10/2026
GraphRAG 1.3.0
The Components service and MCP have little in common:
Components Service:
A stateless intermediary. Your workflow passes it a query string; it retrieves the embedding vector from your provider (OpenAI/Bedrock), executes a raw nearest-neighbor (kNN) search against your vector store, and returns matching document payloads. It is designed for straightforward retrieval—finding documents semantically closest to the query.
MCP (Model Context Protocol):
An open standard developed by Anthropic. Rather than acting as a passive endpoint your workflow calls, MCP exposes your database directly to an LLM as a set of tools. The model can inspect your GraphDB schema, fetch semantic relationship patterns, and discover data graph nodes on the fly guided by its own reasoning.
Vector search is also available as a separate MCP tool that the AI agent can invoke independently. This design is particularly useful for multi-hop scenarios, where answering a question requires combining data from multiple sources through successive tool calls.
For example, the agent might first execute a SPARQL query to retrieve structured data from the graph. Based on that response, it may then perform a vector search to gather additional information that exists in the documents but is not represented in the graph.
The decision of when to call the vector search tool rests with the AI agent and is influenced by both the underlying LLM model and the system prompt. For multi-hop scenarios, explicitly describing these patterns in the system prompt may help guide the agent toward the correct tool selection, though this approach does not guarantee the desired behavior.