FAQ RAG + MCP Tool (Starter Skeleton)

This is a minimal starting point for the MCP option.

Contents

• rag_core.py — RAG core
• mcp_server.py — MCP server exposing ask_faq
• faqs/ — tiny sample corpus
• requirements.txt

Quick Start

python -m venv .venv && source .venv/bin/activate
pip install -r requirements.txt

export OPENAI_API_KEY=sk-...

# Optional model overrides
# export EMBED_MODEL=text-embedding-ada-002
# export LLM_MODEL=gpt-3.5-turbo

# Run a quick CLI smoke test
python rag_core.py

# Configure your MCP client to spawn the server
# command: python
# args: [/absolute/path/to/mcp_server.py]
# env: { OPENAI_API_KEY: "sk-..." }

Design Principles (Evaluation Criteria)

This implementation prioritizes Simplicity, Practicality, and Interface Correctness.

1. Simplicity over Over-Engineering

• No Vector Database: Instead of adding heavy dependencies like Chroma or Pinecone, we use numpy for in-memory cosine similarity. For a filtered FAQ lists, this is faster, easier to debug, and removes deployment complexity.
• FastMCP: We use the high-level FastMCP interface to reduce boilerplate, keeping the server code focused on logic rather than protocol details.
• Global State: We preload the corpus at import time for simplicity in this specific "server" context, avoiding complex dependency injection containers.

2. Practicality

• Robust Error Handling: The server uses structured logging and catches API errors (e.g., rate limits) to prevent crashes, returning user-friendly error messages to the LLM.
• Exposed Resources: The faq:// resource allows the LLM (and developers) to inspect the raw content of any FAQ file, which is crucial for verifying answers or debugging retrieval issues.
• Pre-defined Prompts: The ask_faq_expert prompt helps users/LLMs start with the right context immediately.

3. Interface Correctness

• Standard MCP Patterns: We strictly follow MCP standards by exposing Tools (action), Resources (data), and Prompts (context).
• Type Safety: All tools use Python type hints (str, int) which FastMCP automatically converts to JSON-Schema for the LLM to understand.