Connecting MCP to Verified Knowledge Bases (Sanity Context vs. Naive RAG)

Why Naive Vector Chunking Breaks Technical Documentation

Splitting a tutorial into arbitrary 500-token chunks separates TypeScript interfaces from the prose explaining them. Structured Content Lake queries preserve relationships.

export const AGENT_COURSE_QUERY = `*[_type == "course" && coalesce(archived, false) != true]{
  title,
  "slug": slug.current,
  excerpt,
  "lessons": sections[].lesson[]->{ title, "slug": slug.current, excerpt }
}`;

Production Architecture Takeaway

Query structured documents by schema and reference graph first, then use embeddings strictly for semantic ranking.