Production RAG pipeline + custom MCP

Document ingestion → vector DB → grounded Claude service, with caching, RAGAS evals and a remote MCP exposing internal tools to Claude Desktop.

Engineering · Stack: Python, MCP SDK, Claude API. Outcome: <800ms p95, ship-ready to enterprise

Overview

A production-grade retrieval service that grounds Claude in your documents, with the evals and tooling needed to ship it to an enterprise.

The problem

Demos are easy; a RAG system that is fast, accurate, evaluated and maintainable in production is not.

How it works

• Documents are chunked, embedded and stored in a vector database.
• Queries retrieve and rerank the most relevant context.
• Claude answers grounded in that context, with citations.
• A RAGAS evaluation suite gates every change before deploy.

What it does

• End-to-end ingestion to vector-store pipeline
• Prompt and response caching for latency and cost
• RAGAS evaluation harness wired into CI
• Remote MCP server exposing internal tools to Claude Desktop

Architecture

Python services, a managed vector database, the MCP SDK for the tool server, and the Claude API for generation.

Results

• Under 800ms p95 response time
• Eval-gated deploys catch regressions early
• Ship-ready for enterprise security review