Post

Instead of letting an agent guess which vector chunk matters, start with two things in the config: a sharp dataset purpose and a focused GROQ projection. That produces a clean outline for direct entry-path reads via Context MCP. See it in the reply 👇

28/10/26, 5:15 pm1 repost1 like

Alex Patterson

@codercat.dev

Writing a clear dataset purpose sounds trivial until you watch an agent skip three unnecessary tool calls because the outline told it right away where the answer lives.

CodingCat.dev

@codingcat.dev

See the dataset configuration and GROQ projection workflow in action: https://youtu.be/0gsqAIGxXzc Step-by-step code and config in the post: https://codingcat.dev/post/sanity-knowledge-bases