Skip to main content
Rate limits help protect your applications from misuse and maintain the health of our shared serverless infrastructure. They vary by pricing plan and apply to serverless indexes only. Request-per-second limits are enforced per namespace or per index, as noted in the Data operation throughput limits table, in addition to your read and write unit limits, so a request fails if it exceeds any applicable limit. When you exceed a limit, the request returns a 429 - TOO_MANY_REQUESTS and the error message names the limit, scope, and value. Most rate limits can be adjusted upon request. If you need higher limits, contact Support with details about your use case.
Indexes built on Dedicated Read Nodes aren’t subject to read unit limits for query, fetch, list, and full-text search operations. For sizing and capacity planning guidance, see the Dedicated Read Nodes guide.

Monthly usage limits

Read units and write units measure resource consumption. To check your current usage, see Download a usage report.
From November 4 through December 31, 2026, Starter orgs get 10,000,000 embedding tokens per month per model, and Builder orgs get 20,000,000. Allowances return to the values in the table on January 1, 2027.
Monthly rerank request limits vary by model: Reaching a monthly usage limit returns a 429 - TOO_MANY_REQUESTS. Upgrade your plan to continue.

Data operation throughput limits

The per-namespace and per-index query read unit limits are enforced separately. At the defaults they’re equal, so the namespace limit only matters once the index limit is raised. When you ask Support for a higher read limit, ask for both, or each namespace stays capped at its current limit. Exceeding a per-second throughput limit returns a 429 - TOO_MANY_REQUESTS. Implement retry logic with exponential backoff, pace your requests, consider Dedicated Read Nodes (which aren’t subject to per-second read limits) for high-throughput reads, or contact Support to raise a limit.

Model throughput limits

Per-minute token limits for embedding models: Per-minute request limits for reranking models: Reaching a per-minute model limit returns a 429 - TOO_MANY_REQUESTS. Upgrade your plan to increase it, or implement retry logic with exponential backoff.

Inference request limits

These limits apply per project across all models. Every inference request counts toward them, including requests that a per-minute token limit then rejects. Exceeding the per-second or per-minute limit returns a 429 - TOO_MANY_REQUESTS. Implement retry logic with exponential backoff to handle it.
The message names the per-second or per-minute limit, whichever you exceeded.