SQL‑AI Samples MCP Server

The SQL‑AI Samples server turns an Azure SQL Database into a first‑class AI data source that Claude and other LLMs can query, embed, and analyze on the fly. By exposing a collection of pre‑built prompts, vectorization routines, and retrieval pipelines through the Model Context Protocol, it eliminates the need for developers to write custom connectors or manage embeddings manually. Instead, the server provides a unified interface that blends structured SQL data with modern AI workflows such as semantic search, conversational agents, and content moderation.

At its core, the server solves a common pain point: how to make relational data conversationally accessible. Traditional SQL queries are declarative but opaque to end users; conversely, generative models can understand natural language but lack direct access to structured datasets. This MCP server bridges that gap by letting the model issue high‑level requests—“find me the top‑selling products in Q3” or “rank customers by churn risk”—and then translating them into efficient T‑SQL statements that run against the underlying database. The results are streamed back as JSON or plain text, ready for downstream LLM processing.

Key capabilities include:

• Prompt‑driven SQL generation: Developers supply a natural‑language prompt, and the server invokes an LLM to produce safe, parameterized T‑SQL code that respects schema constraints. This reduces the risk of injection and promotes reusable query patterns.
• Vector embedding integration: Built‑in hooks to Azure Cognitive Search and Azure OpenAI enable the server to compute embeddings for text columns or custom fields, store them in auxiliary tables, and perform cosine‑similarity searches directly from the LLM. This makes semantic search over structured data trivial.
• Retrieval‑augmented generation (RAG): The server can fetch relevant rows, embed them, and feed the context back to the LLM, allowing for highly contextual answers that reference actual database contents.
• Content moderation and policy enforcement: Sample workflows demonstrate how to run user‑generated content through Azure’s moderation services before it reaches the database, ensuring compliance with organizational policies.

Real‑world scenarios that benefit from this server are plentiful. An e‑commerce platform can build a conversational chatbot that recommends products by querying sales and inventory tables in real time. A financial analyst can ask for a risk assessment of a portfolio, with the server pulling in transaction histories and performing on‑the‑fly calculations. In research settings, scholars can query a bibliographic database for papers similar to a given abstract using vector similarity, all through natural language.

Integration into AI pipelines is straightforward. The MCP server exposes endpoints for resources (the database tables), tools (SQL execution, embedding generation), and prompts (pre‑defined question templates). A Claude client can invoke these tools in a single turn, receiving structured results that feed into subsequent LLM prompts. This composability means developers can chain together complex workflows—e.g., retrieve similar articles, summarize them, and generate a report—all without leaving the LLM context.

What sets this MCP server apart is its end‑to‑end sample ecosystem. The repository bundles notebooks and end‑to‑end demos that cover everything from simple product recommendation to sophisticated vector search with FAISS or IVFFlat indexes. By providing ready‑made templates and a clear mapping between natural language, SQL, and embeddings, it lowers the barrier for developers to embed relational data into AI applications.