Vectra MCP Server

A Model Context Protocol (MCP) server for interacting with a Vectra knowledge base.

This TypeScript-based MCP server provides tools to manage and query a Vectra instance, enabling integration with MCP-compatible clients. It interacts with a backend Vectra API (presumably running separately).

Features

Tools

This server exposes the following tools for interacting with Vectra:

• create_collection: Create a new Vectra collection.
  • Input: name (string, required), description (string, optional)
• list_collections: List existing Vectra collections.
  • Input: None
• embed_texts: Embeds multiple text items in batch into Vectra.
  • Input: items (array of objects with text (required) and optional metadata), collectionId (string, optional)
• embed_files: Reads multiple local files and embeds their content into Vectra.
  • Input: sources (array of local file paths, required), collectionId (string, optional), metadata (object, optional - applies to all items)
• add_file_to_collection: Add an already embedded file (referenced by its ID) to a specific Vectra collection.
  • Input: collectionId (string, required), fileId (string, required)
• list_files_in_collection: List files within a specific Vectra collection.
  • Input: collectionId (string, required)
• query_collection: Query the knowledge base within a specific Vectra collection.
  • Note: This tool always uses hybrid search (vector + keyword) and enables graph search enhancement by default.
  • Input: collectionId (string, required), queryText (string, required), limit (number, optional), maxDistance (number, optional), graphDepth (number, optional), graphRelationshipTypes (array of strings, optional), includeMetadataFilters (array of objects, optional), excludeMetadataFilters (array of objects, optional)
• delete_file: Delete a file and its associated embeddings from Vectra.
  • Input: fileId (string, required)
• get_arangodb_node: Fetch a specific node directly from the underlying ArangoDB database by its key.
  • Input: nodeKey (string, required - e.g., chunk_xyz or doc_abc)

(Refer to src/tools.ts for detailed input schemas)

Development

Install dependencies:

npm install

Build the server:

npm run build

Run the server (listens on stdio):

node build/index.js

For development with auto-rebuild:

npm run watch