Add cloudberry-mcp-server for integration with LLM apps

[yangs16]

Proposers

Shengwen Yang

Proposal Status

Under Discussion

Abstract

This proposal introduces cloudberry-mcp-server, a new sub-project within Apache Cloudberry that implements the Model Context Protocol (MCP) specification. This server will bridge Cloudberry's powerful database capabilities with modern LLM applications, enabling natural language interactions with structured data and enterprise knowledge bases. End users will be able to list tables, query databases using natural language (automatically converted to SQL), visualize query results, and leverage Cloudberry's robust embedding and full-text search capabilities for powerful RAG (Retrieval Augmented Generation) applications. Built on the FastMCP foundation, this sub-project will transform how organizations build and deploy LLM applications that require enterprise data access while maintaining Cloudberry's performance and scalability.

Motivation

As Large Language Models (LLMs) become increasingly integrated into enterprise workflows, there's growing demand for natural language interfaces to both structured data and proprietary knowledge bases. While Cloudberry provides robust database and search capabilities, it currently lacks standardized integration with LLM applications that could simplify RAG implementations. The Model Context Protocol (MCP) offers an ideal standard for this integration.
Key motivations for this sub-project include:

• Natural Language Data Access: Enable users to interact with Cloudberry databases through natural language, without requiring SQL expertise or database knowledge.
• Enterprise Knowledge Base Integration: Allow organizations to build proprietary knowledge bases in Cloudberry using both embedding and full-text search capabilities, making RAG LLM applications significantly easier to deploy.
• Simplified RAG Implementation: Provide a turnkey solution for organizations looking to augment LLMs with their proprietary data and knowledge bases stored in Cloudberry.
• Multi-Modal Search Capabilities: Leverage Cloudberry's combined embedding search and full-text search capabilities to provide more accurate and contextually relevant responses through LLM applications.
• Reduced Technical Barriers: Lower the implementation complexity for creating enterprise-grade LLM applications with accurate, up-to-date information retrieval.

Implementation

1. MCP Server Implementation

• Built using FastAPI, leveraging the FastMCP package
• Implementation of all required MCP endpoints for context retrieval
• Extensions for database metadata exploration (table listing)

2. Retrieval Augmented Generation

• Hybrid search capabilities combining embedding and full-text search for optimal retrieval
• Configuration interface for managing knowledge bases stored in pre-built Cloudberry tables
• Advanced relevance scoring with adjustable weighting between semantic and keyword matches

3. Natural Language to SQL Translation

• Integration with LLM-powered SQL generation capabilities
• Query validation and safety checks before execution

Rollout/Adoption Plan

No response

Are you willing to submit a PR?

• Yes I am willing to submit a PR!

[tuhaihe]

Thank you for proposing this idea!

Integrating Cloudberry with LLM applications via the MCP is a cool direction that aligns well with current trends and potential user needs. Give +1 to this proposal. If we can proceed successfully, I suggest we create a separate repo to maintain this project (maybe call cloudberry-mcp-server, I can help with this)

Let me know if I can help in any way! Also, let's wait for more voices from the community.

[yjhjstz]

Can we impl mcp-server also be an extension ? not just write by python ?

[yangs16]

Not sure about this. MCP-server is expected to be integrated with LLM clients/agents. If it's implemented as a Cloudberry extension, how will the LLM clients/agents interact with it?

[yjhjstz]

we can write extension using pgrx(which Build Postgres Extensions with Rust) or c/c++ .

[reshke]

I dont see why this cannot be an extension actually

[yjhjstz]

refer to https://github.com/HenkDz/postgresql-mcp-server, most mcp written by TypeScript, so to be cloudberry extension, kind of weird.

[tuhaihe]

Hey, how's the progress on this proposal? Since developing it as an extension with PGRX or C/C++ is somewhat challenging, we can instead build this feature on the FastMCP foundation, as outlined in the proposal.

Also, we could include this feature in the src/tools/ directory, since we now have too many sub-repositories. Love to hear your feedback again. cc @yjhjstz @yangs16

[yjhjstz]

Maybe a single alone repo is better .

[tuhaihe]

Update: Discussed with @my-ship-it and @yangs16. From the perspective of market promotion and user usability of this feature, we now prefer to place it in the top-level directory and create a new folder similar to mcp-server. This feature can reuse the Python environment used by the database system.

For the src/tools/, it's used for storing the database core codes. And for a separate repo, now we have lots of repos, but lost maintenance. For cloudberry-devops-release and cloudberry-bootcamp, will also be reorganized into the main repo for better organization. Hope it makes sense.