Similarity search is a leaky abstraction for structured knowledge. Instead of hoping top-k vector chunks have the right context, give your agent a purpose-steered outline via `initial_context` and read exact Markdown paths. Architecture in the reply 👇

Naive vector RAG chunks raw documents and forces your LLM to resolve contradictions at query time. Hardcoding Markdown in system prompts breaks when content updates. Ahead-of-time Knowledge Base compilation fixes both. Full walkthrough in the reply 👇