Modular RAG MCP Server

Skill by ara.so — MCP Skills collection.

Expert skill for deploying, configuring, and extending the Modular RAG MCP Server — a pluggable, observable RAG (Retrieval-Augmented Generation) system that exposes tools via Model Context Protocol for AI assistants like Claude Desktop and GitHub Copilot.

What This Project Does

The Modular RAG MCP Server is a complete RAG pipeline featuring:

• Ingestion Pipeline: PDF → Markdown → Chunking → Embedding → Vector Store (with multimodal image captioning)
• Hybrid Search: Dense vectors (semantic) + Sparse BM25 (exact match) + RRF fusion + optional reranking
• MCP Protocol: Standard MCP server exposing query_knowledge_hub, list_collections, get_document_summary tools
• Dashboard: Streamlit-based management UI with 6 pages (overview, data browser, ingestion tracking, query tracking, evaluation)
• Evaluation Framework: Ragas + custom metrics for regression testing
• Full Observability: White-box tracing of ingestion and query pipelines

Key Architecture: Every core component (LLM, Embedding, Reranker, Splitter, VectorStore, Evaluator) is pluggable via abstract interfaces. Switch backends through configuration without code changes.