Guides · Developer
Developer guide
Patterns we've seen work well in production.
Caching
Cache identical Search queries for the duration the freshness window allows. For `freshness=day`, a 1-hour TTL is usually safe; for evergreen queries, cache aggressively.
Rate limits
Use exponential backoff with jitter on 429s. Concurrency caps are per-key.
Citations
- Always show source URLs to end-users for grounded answers.
- Render snippets exactly as returned — they map back to the linked passage.
Cost control
- Prefer Search + selective Contents over deep Research for high-volume traffic.
- Use `num_web_results` to cap the bytes you pay to process.