Module 4 · Lecture 4.6
Context: Layers, Notes & When NOT to Use Them
The context stack from deterministic to probabilistic, and how to decide which layer a piece of knowledge belongs in.
Veezoo can hold business context in several layers, from fully deterministic to probabilistic: SQL and VQL definitions with names and synonyms in the Knowledge Graph, then descriptions, ground truths, notes, and static notes.
This lecture walks the stack and then shows a live demo, taking three static notes and pushing each one as far toward the deterministic end as it can go.
Key points
- The context stack runs from deterministic to probabilistic: VQL definitions and names or synonyms, then descriptions, then ground truths, then notes, then static notes.
- Push every piece of context as far toward the deterministic end of that stack as it can go.
- Static notes are always on, included in every prompt with every question, which makes them useful for documentation-style context and expensive to keep.
- Notes are free-text business context with a name, synonyms, and a description, and they load only when they are relevant to the question.
- Ground truths are question and VQL pairs that act as templates, biasing Veezoo toward the preferred answer for complex calculations.
- Where filters are static or a KPI has a clear definition, a VQL definition beats a note, because it is unambiguous, can be suggested as a filter in chat, and other measures can build on it.
- In the demo: brand guidelines become a note, a complex cross-fact-table calculation becomes a ground truth, and imported BI metrics become measures that Veezoo then tests.
Check your understanding
What is the difference between a note and a static note?
Static notes are always-on context sent with every question, while notes are matched by name, synonyms, and description.
A static note reads "Active customer = customer with at least one order in the last 90 days. Use this for all activity KPIs." Where does this context belong?
The filter is static and the definition is clear, so a VQL definition is unambiguous and other measures can build on it.
What is a ground truth in Veezoo?
Ground truths pair a question with its preferred VQL, which biases Veezoo toward that answer for complex calculations.