Atlas Vector Search Automated Embedding rate limits
Parent: Atlas Pricing Programs, Cost Governance and TCO · Published reference · snapshot 2026-10-21 · skill ranked-3d355de6
↓ Facts as markdownall context files
Source-anchored research for Atlas Vector Search Automated Embedding rate limits. Run status: COMPLETED.
These notes link each claim to its source. Some citations identify archived research reports that are no longer hosted on this site. A published reference means the content is available; it does not certify independent review or accuracy.Read the editorial policy and follow the sources before relying on a claim.
Problems, failure modes and limitations
- As published on 2026-10-19, query embeddings on M0 clusters without a payment method have a limit of 3 RPM and 2,000 TPM for each supported model. Query limits on M0 with a payment method, Flex, and Dedicated clusters are 2,000 RPM and model-specific TPM: 3 million for voyage-4-large and both code models, 8 million for voyage-4, and 16 million for voyage-4-lite. The insert/update limits table lists 2,000 RPM and the same model-specific TPM values. (high confidence) (verified as of 2026-10-19) [source]
- AutoEmbed moves document and query embedding generation into Atlas-managed operations, reducing the need for an application-owned embedding pipeline. Taken together, the independent overview and MongoDB billing/rate-limit docs indicate that this shifts operational work but does not remove per-token charges or capacity planning. (medium confidence) (verified as of 2026-10-19) [source]
- MongoDB's overview says over-limit queries return a rate-limit error, while over-limit document updates are queued and embedding generation can lag even as the application continues operating. The models page separately describes query and index-update limits. (medium confidence) (verified as of 2026-10-14) [source]
- Atlas-created Voyage model API keys expose their Automated Embedding rate limits in Atlas, while MongoDB directs users with keys created directly at Voyage AI to Voyage's rate-limit documentation. Voyage's published direct API tiers are organization/project scoped, so those provider limits should not be assumed to describe Atlas-managed cluster quotas. (high confidence) (verified as of 2026-10-21) [source]
- For large first-time builds, MongoDB says Automated Embedding uses a separate inference mechanism that is not bound by standard rate limits and can scale bursts up to available GPU capacity. The overview also says initial-sync duration depends partly on available rate-limit quota, so the exception is not a documented completion-time guarantee. (medium confidence) (verified as of 2026-10-19) [source]
- MongoDB directs users with Voyage AI keys created directly through Voyage to Voyage AI's separate rate-limit documentation; Voyage describes its API limits at the organization level, with project limits set beneath them. Atlas-managed Automated Embedding and direct Voyage API capacity should therefore be budgeted as separate quota systems, despite sharing model names. (medium confidence) (verified as of 2026-08-26) [source]
Facts and statements
- MongoDB enforces Automated Embedding limits in requests per minute (RPM) and tokens per minute (TPM) at the cluster level, with limits shared by all Automated Embedding indexes on that cluster. It separates limits for query embedding, first-time index builds, and insert/update re-embedding; MongoDB says index-build processing is isolated from real-time query traffic. (high confidence) (verified as of 2026-10-19) [source]
- MongoDB says free-cluster users can increase Automated Embedding limits by adding a payment method to move to a paid tier. Paid-tier limits increase automatically as usage grows; teams that need still higher rates must contact MongoDB Support or their account team. (medium confidence) (verified as of 2026-10-19) [source]
- Atlas bills tokens used during initial index sync, document inserts and updates, and queries. Atlas grants one one-time allocation of 200 million free tokens per model at the organization level, shared across projects and clusters, and the tokens do not refresh. This is a cost allowance; it does not replace the separate RPM/TPM throughput quotas. (high confidence) (verified as of 2026-10-19) [source]
- The Hindsight Guild README reports one M0 free-tier public-preview run where bulk inserts exhausted Automated Embedding quota and temporarily caused query-time embedding failures, with about 75 minutes to recover. This is one project's report, not a general SLA or current cap. (low confidence) (verified as of 2026-10-21) [source]
- Atlas usage pages and metrics break out embedding requests and tokens for indexing and query operations. MongoDB says a free cluster can move to a paid tier by adding a payment method; paid-tier limits rise automatically with usage, and MongoDB Support is the path for further increases. (high confidence) (verified as of 2026-10-19) [source]
Children
- No children recorded.