Teach the vision model your world before it looks

Give the vision and reasoning models a few sentences of your own vocabulary before indexing, and everything indexed afterwards is better for it.

Set this as GroundSlate's vision context before it indexes anything else: everything it is about to look at is a Pakistani street food channel; deep fried chicken at a Kohat bazaar stall, the cook is Ruzi Khan, the huge flat pan is a kadai, the yellow coating is spiced chickpea flour. Give the reasoning model the same vocabulary, so summaries and event names use it too. Then index /Volumes/Cards/2026-08, and when the first clips finish, show me their summaries so I can see whether the vocabulary took.

Paste it into any MCP client

It shows up in search, shot descriptions, summaries and event names.
If your library reads as generic, a missing context is usually why, not the model.