MCP Server For Vertex AI Search

What is MCP Server For Vertex AI Search about?

A server that leverages Vertex AI Gemini with grounding to search across one or multiple Vertex AI Datastores, returning results that are contextually anchored in your private data.

How to use MCP Server For Vertex AI Search?

1. Clone & Set Up

   git clone https://github.com/ubie-oss/mcp-vertexai-search.git
   cd mcp-vertexai-search
   uv venv
   uv sync --all-extras
   uv run mcp-vertexai-search
   

2. Or install the package directly

   pip install git+https://github.com/ubie-oss/mcp-vertexai-search.git
   mcp-vertexai-search --help
   

3. Configure
   • Copy config.yml.template to config.yml.
   • Fill in server name, model settings (model name, project ID, location, service‑account impersonation, generate‑content config) and list of Vertex AI Datastores (project ID, location, datastore ID, tool name, description).
4. Run the server (choose transport)

   uv run mcp-vertexai-search serve \
       --config config.yml \
       --transport <stdio|sse>
   

5. Test a search without the server

   uv run mcp-vertexai-search search \
       --config config.yml \
       --query "your query"
   

Key Features

• Grounded Search: Uses Gemini’s grounding capability to tie responses to your own Vertex AI Datastores.
• Multi‑Datastore Support: Integrate one or many Datastores simultaneously.
• Flexible Transport: Supports Server‑Sent Events (SSE) and standard I/O transports.
• Config‑Driven: All settings (model, project, datastores) are managed via a simple YAML file.
• Docker Ready: Dockerfile provided for containerized deployment.

Use Cases

• Enterprise knowledge bases where search must respect proprietary data.
• Internal documentation portals powered by LLMs with accurate references.
• Customer support tools that retrieve up‑to‑date information from private datasets.
• Research environments needing grounded retrieval across multiple data sources.

FAQ

Q: Do I need a Google Cloud account? A: Yes, you need access to Vertex AI models and Datastores within a GCP project.

Q: Is the package available on PyPI? A: Not yet; install directly from the GitHub repository.

Q: Which model should I use? A: Any Gemini model that supports the generate‑content API; specify the model name in config.yml.

Q: Can I run this in production? A: Yes, using the provided Dockerfile or by deploying the server with your preferred orchestration tool.

Q: How do I add multiple Datastores? A: List each datastore under the data_stores section in config.yml with its project ID, location, ID, tool name, and description.

MCP Server for Vertex AI Search

This is a MCP server to search documents using Vertex AI.

Architecture

This solution uses Gemini with Vertex AI grounding to search documents using your private data. Grounding improves the quality of search results by grounding Gemini's responses in your data stored in Vertex AI Datastore. We can integrate one or multiple Vertex AI data stores to the MCP server. For more details on grounding, refer to Vertex AI Grounding Documentation.

How to use

There are two ways to use this MCP server. If you want to run this on Docker, the first approach would be good as Dockerfile is provided in the project.

1. Clone the repository

# Clone the repository
git clone git@github.com:ubie-oss/mcp-vertexai-search.git

# Create a virtual environment
uv venv
# Install the dependencies
uv sync --all-extras

# Check the command
uv run mcp-vertexai-search

Install the python package

The package isn't published to PyPI yet, but we can install it from the repository. We need a config file derives from config.yml.template to run the MCP server, because the python package doesn't include the config template. Please refer to Appendix A: Config file for the details of the config file.

# Install the package
pip install git+https://github.com/ubie-oss/mcp-vertexai-search.git

# Check the command
mcp-vertexai-search --help

Development

Prerequisites

• uv
• Vertex AI data store
  • Please look into the official documentation about data stores for more information

Set up Local Environment

# Optional: Install uv
python -m pip install -r requirements.setup.txt

# Create a virtual environment
uv venv
uv sync --all-extras

Run the MCP server

This supports two transports for SSE (Server-Sent Events) and stdio (Standard Input Output). We can control the transport by setting the --transport flag.

We can configure the MCP server with a YAML file. config.yml.template is a template for the config file. Please modify the config file to fit your needs.

uv run mcp-vertexai-search serve \
    --config config.yml \
    --transport <stdio|sse>

Test the Vertex AI Search

We can test the Vertex AI Search by using the mcp-vertexai-search search command without the MCP server.

uv run mcp-vertexai-search search \
    --config config.yml \
    --query <your-query>

Appendix A: Config file

config.yml.template is a template for the config file.

• server
  • server.name: The name of the MCP server
• model
  • model.model_name: The name of the Vertex AI model
  • model.project_id: The project ID of the Vertex AI model
  • model.location: The location of the model (e.g. us-central1)
  • model.impersonate_service_account: The service account to impersonate
  • model.generate_content_config: The configuration for the generate content API
• data_stores: The list of Vertex AI data stores
  • data_stores.project_id: The project ID of the Vertex AI data store
  • data_stores.location: The location of the Vertex AI data store (e.g. us)
  • data_stores.datastore_id: The ID of the Vertex AI data store
  • data_stores.tool_name: The name of the tool
  • data_stores.description: The description of the Vertex AI data store