Skip to main content
Mixpeek enforces limits at multiple layers—API rate limits, credit consumption, retriever budgets, and infrastructure quotas. Understanding each layer helps you avoid 429 errors and unexpected costs.

​
API Rate Limits

  • Per-route counters backed by Redis (SlowAPI).
  • Tier defaults (requests/minute):
    • Free: 10 (general), 10 (retriever execute, uploads)
    • Pro: 60
    • Enterprise: configurable (contact support)
  • Responses include details.api_name to identify which bucket tripped the limit.
Recommendations
  1. Implement exponential backoff with jitter (1s, 2s, 4s, …).
  2. Batch uploads using /objects/batch and process via batches instead of hot loops.
  3. Enable retriever caching to reduce duplicate execute calls.

​
Credits & Usage

Credits measure compute-intensive activity (extractions, retriever execution, LLM stages). Track usage with analytics endpoints: 
curl -sS -X GET "$MP_API_URL/v1/analytics/usage/summary?start_date=2025-10-01&end_date=2025-10-31" \
  -H "Authorization: Bearer $MP_API_KEY" \
  -H "X-Namespace: $MP_NAMESPACE"
Monitor retriever cost with the budget section in execution responses: 
"budget": {
  "credits_used": 18.75,
  "credits_limit": 100,
  "time_elapsed_ms": 5250
}
Set budget_limits on retrievers to cap runtime and credit spend.

​
Task Backlog Limits

  • Celery queues and Redis memory impose soft limits on concurrent tasks.
  • Monitor queue depth; auto-scale workers or throttle submissions if tasks pile up.
  • Use batch sizes that match available Ray worker capacity.

​
Ray Serve & Model Quotas

  • Each model deployment has min_replicas/max_replicas and per-request concurrency targets.
  • For LLM providers (OpenAI, Anthropic, etc.), respect provider-side rate limits; Mixpeek surfaces provider errors directly.

​
Storage Considerations

ComponentLimitMitigation
S3Object count/sizeEnable lifecycle policies, compress artifacts
MongoDBDocument size (16 MB)Keep metadata lean; store large payloads in S3
QdrantMemoryShard by namespace, prune stale documents, upgrade memory

​
Handling 429 Too Many Requests

  1. Inspect error.details.api_name.
  2. Back off and retry later (respect Retry-After headers if supplied).
  3. Reduce concurrency from your client or cache repeated reads.
  4. Consider upgrading your tier if usage is sustained.

​
Monitoring Usage

  • Log API responses for 429, 403, and NotEnabledError to detect limit issues.
  • Build dashboards using analytics endpoints (retriever performance, cache hit rate, slow queries) to see when workloads approach limits.
  • Use webhook events to trigger scaling actions (e.g., new batch submitted → scale Ray workers).

​
When to Contact Support

  • You need higher rate limits or custom credit pools.
  • You plan to run batch workloads that exceed current quotas.
  • Provider-specific limits (LLM APIs) block your workflow even after retries.
Include namespace, organization, relevant request IDs, and desired limits when reaching out.

​
References