Skip to main content

​
BigQuery smart prompting framework on Julius

​
1. Always include time boundaries

Why this matters: Tables can contain years of data. Without time limits, Julius might scan everything.
Costly approach: “What are our sales trends?”
Cost-effective approach: “Show me daily sales trends for the past 3 months”
Why it’s better: The first query forces Julius to scan the entire sales history (potentially years of data). The second query uses date partitioning to scan only 3 months of data - potentially reducing processing by 80-90%. Pro tip: Start with recent data for exploration, then expand timeframes only when needed.

​
2. Be specific about what you need

Why this matters: Julius will include relevant columns and metrics based on your request. Vague requests lead to broader data pulls.
Costly approach: “Analyze our customer data”
Cost-effective approach: “Show me customer purchase frequency and average order value for active customers in Q4”
Why it’s better: The first query might pull every customer attribute across all time periods. The second targets specific metrics (purchase frequency, AOV) for a defined segment (active customers) in a specific period (Q4). This could reduce data processing by 70-95%.

​
3. Use smart sampling for exploration

Why this matters: When you’re exploring patterns or testing hypotheses, perfect precision often isn’t necessary.
Costly approach: “What patterns do you see in user behavior?”
Cost-effective approach: “Analyze login patterns and session duration using a 20% sample of users from the past month”
Why it’s better: The first query processes every user interaction ever recorded. The second processes only 20% of recent data, which is typically sufficient for pattern recognition and costs 80% less while still providing statistically valid insights.

​
4. Layer your analysis

Why this matters: Start broad with cost-effective queries, then drill down based on what you discover.
Costly approach: “Give me a complete breakdown of all our metrics by every possible dimension”
Cost-effective approach: “Show me our top 5 product categories by revenue this quarter, then I’ll dive deeper into the most interesting one”
Why it’s better: The first approach processes massive amounts of data upfront. The layered approach finds the most important areas first (small data scan), then focuses detailed analysis only where it matters most.

​
5. Reference existing summaries when available

Why this matters: We have pre-computed summary tables and views that contain aggregated data.
Costly approach: “What’s our monthly revenue growth?”
Cost-effective approach: “Using our monthly_revenue_summary table, show me growth rates for the past year”
Why it’s better: The first query might aggregate daily transaction data from scratch. The second uses pre-computed monthly summaries, processing 12 rows instead of potentially millions.

​
Quick check: The smart prompting checklist

Before submitting your query, check:
  • ✅ Time Range: Did you specify when? (past month, Q4, last 90 days)
  • ✅ Scope: Are you asking for specific metrics or segments?
  • ✅ Purpose: Is this exploration (use sampling) or precision analysis?
  • ✅ Building Blocks: Can you start with a summary or build on previous work?
Reach out to team@julius.ai for support or to ask questions not answered in our documentation.