Query response time and quotas

更新时间:
复制 MD 格式

This topic describes how to optimize performance and plan resource usage by understanding factors that affect response time (such as feature configurations, data center locations, and query length) and quota management rules (such as QPS specifications and throttling mechanisms).

Search response time

The search API response time is affected by the following factors:

  1. Search rerank feature (enabled by default): After retrieval, a rerank model performs semantic reranking. This improves relevance mean Average Precision (mAP) by about 5% but adds about +140 ms to the response time. To disable this feature, contact us.

  2. timeRange auto-rewrite feature (disabled by default): This feature automatically rewrites the `timeRange` parameter based on the query's semantics. It only adjusts the time range of the search and adds about +300 ms to the response time.

  3. Data center location: The service's data center is in Zhangjiakou. If your data center is far from Zhangjiakou, latency can increase by tens of milliseconds. If your service is deployed in an Alibaba Cloud VPC, you can use a VPC endpoint to call our service (General Search). This helps reduce latency and improve stability.

  4. Query length: Longer queries cause a slight increase in search latency.

    • The following search response times are for reference only. These times do not include latency from reranking, `timeRange` rewriting, or server-side statistics. Actual response times may vary slightly depending on the query:

    Average RT

    950 ms

    P90

    1200+ ms

    P99

    1900+ ms

    Note

    You can check the searchInformation.searchTime parameter in the API response to view the server-side processing time and analyze network latency.

Search quota

The service is available in 30 queries per second (QPS) and 60 QPS specifications. Prices vary because each specification comes with a different level of resource assurance. To reduce throttling (error code 429/Retrieval.Throttling.User) caused by traffic spikes, we calculate QPS over a 3 second time window. Therefore, a 30 QPS limit is equivalent to 90 queries per 3 seconds.