Rules vs example queries: what context helps your analytics agent the most?

Why are examples so often recommended for LLM context? The response is simple: they reduce the amount of work an LLM has to do to give a similar response.

A not too simple example: for the question "What is the month-over-month revenue growth rate over the last 12 months?", it's very different if a model sees:
```

Table JAFFLE.ORDERS

SUBTOTAL (DOUBLE): Order value before sales tax, in USD.
The basis for revenue and order-value figures.
...

Example queries

What is the month-over-month order-count growth rate (%) for each month?

WITH monthly AS (
SELECT date_trunc('month', ORDERED_AT) AS month, COUNT(*) AS orders
FROM ORDERS GROUP BY 1
)
SELECT month, orders,
100.0 * (orders - LAG(orders) OVER (ORDER BY month))
/ LAG(orders) OVER (ORDER BY month) AS growth_pct
FROM monthly ORDER BY month
...
or:
...
Metric: revenue
- Expression: SUM("SUBTOTAL")
- Table: JAFFLE.ORDERS
...
```

However, an example library is hard to maintain and to keep consistent, and it can never cover all the questions users will ask. It's easier to just state rules in plain English and let the LLM follow them. So how much performance do they bring exactly? And what are possible alternative ways of presenting the facts required to answer?

This talk will be a presentation of our research on this topic, focused on agentic analytics use cases. I will describe the experiments, with specific numbers on the latest models, and all the mistakes I did while attempting to measure this, so attendees know the traps to avoid. It will conclude on the key decisions and tradeoffs one can make when assembling context, depending on the performance, cost, and maintainability goals.