RAG MCP Integration

Retrieval design (chunking, indexing, hybrid search, reranking), MCP-style tool integration patterns, and the prompt-cache-aware context block. The 'why retrieval first' part of the tuning order.

Interview talking points

• How do you decide chunk size? Trade-off between recall and signal-to-noise; tune via offline harness, not vibes.
• Hybrid (BM25 + dense) vs dense only? Hybrid still wins for keyword-heavy queries; matters for product-catalog retrieval.
• Cohere rerank — where does it fit? As a top-K post-filter when the encoder is good but the dataset is broad.
• MCP tool schema vs Anthropic tool_use vs OpenAI function-calling. The schema is the same shape — that's the interop story.

Files in this folder

Back to the home page.