Skip to main content
Search latency varies with the request and the amount of content being searched. Measure performance with your own library and query patterns before setting client timeouts.

​
Improve request performance

  • Pass collection_id when you know which collection contains the content.
  • Request only the search modes your workflow needs.
  • Keep limit and page sizes as small as your interface allows.
  • Avoid sending image input when a text query is sufficient.
  • Use natural_language=true only for conversational queries that benefit from additional interpretation.
Collection-scoped searches usually examine less content than account-wide searches. This can improve both latency and result relevance.

​
Set client timeouts

Allow for normal variation instead of depending on a single latency estimate. Log meta.request_id and meta.took_ms when available so you can investigate slow requests. Retry transient server errors with exponential backoff and jitter. Do not automatically retry validation or authentication errors.

​
Rate limits

Rate limits control request frequency. They are separate from the response time of an individual request. See Rate limits for current limits and 429 handling.