Getting Started with grafter; A step-by-step guide: install it, run it, read the outputs, and know when it pays off

Description
Grafter takes a raw fact table — an entity, a time column, a measure (e.g. customer, date, qty) — and grafts on the analytical features a senior analyst would compute before forming a hypothesis (trend, volatility, cadence, anomaly, momentum), each with a machine-readable definition. The goal isn't to make a model "smarter." It's a targeted tool: in a cached, repeated-query pipeline on dense data, letting a model read a pre-computed statistic instead of deriving it from raw rows makes analytical queries cheaper and more reliable. This approach gives higher accuracy, Cheap model 28% → 62%, strong model 20% → 76%, and better reasoning for cheaper models. Anyone setting up data for LLMs to read has to use some approach like this. 


Use this:

✅ Deploy grafter when:

• Queries are repeated against a stable context (dashboards, monitoring agents, repeated analytical Q&A).
• The context is prompt-cached (this is what makes it cheaper).
• Per-entity history is dense (~15+ events) so features are stable.
• You want a cheaper model to answer derived-statistic questions reliably.
  
  

⛔ Skip it when:

• One-off queries (no cache reuse — grafting costs more uncached).
• Thin per-entity history (few events) — features become unstable and can hurt accuracy.
• A capable model with a SQL tool already answers correctly and cheaply.