MCP is a solid integration layer — but how does it hold up when it comes to output quality? Often, not as well as you'd like. Here are some practical retrieval patterns, from basic to advanced, that worked well in my experiments:
* Naive: Just plug in plain MCP and hope the LLM gets it right. Sometimes it does. Sometimes you’ll need a miracle.
* Semantic: Add more descriptive field names and extra metadata. It helps — but usually just a bit.
* Templated: Use a structured template and have the LLM fill it out step by step. More effort, but by far the most reliable results.