Reference
Limits
Every numeric limit on gateway requests, Functions and API keys in one place.
Limits apply per workspace unless noted. Contact us if a limit blocks your use case.
Gateway#
| Limit | Value | Notes |
|---|---|---|
| Request deadline | 55 seconds | Covers queueing, retries and provider fallback. Expiry returns 504 request_deadline_exceeded. |
| Idempotency replay window | 120 seconds | Non-streaming requests only. Keys are 1–255 characters, scoped to the API key. Responses over 64 KiB are not replayed. |
| Response cache TTL | Default 3,600 s, maximum 604,800 s (7 days) | Larger cache.ttl_s values are clamped to the maximum. |
| Semantic cache similarity | Default 0.97 | Override with cache.similarity_threshold (0–1). |
| Correlation header IDs | 200 bytes each | X-Seldon-Session-Path allows 512 bytes; X-Seldon-Session-Metadata allows 2 KiB. |
| End-user identifier | 256 characters | Longer user values are truncated for attribution. |
| Provider rate limits | Per provider and model | Listed as rate_limit_rpm, rate_limit_tpm and rate_limit_tpd in GET /v1/models. |
Functions#
| Limit | Value | Notes |
|---|---|---|
| Functions per workspace | 50 | Delete unused functions to free a slot. |
| Request size | 256 KiB | Applies to drafts and to each call. Larger calls return 413 request_too_large. |
| Calls per workflow | Up to 3 | Each call accepts up to 32 messages. |
| Call deadline | 55 seconds | Includes every stage and any fallback to the original workflow. |
| Workspace concurrency | 10 in flight, 5 per second | Excess calls return 429 execution_capacity. |
| Per-call spending reservation | $10 (earlier revisions: $1) | Held while a call runs; you are charged actual usage. |
| Compile cap | Default $10, maximum $100 | Covers preview, search and measurement for one run. |
| Optimization deadline | 2 hours | Includes queue time. Results measured before the deadline remain available. |
| Preview quote | Valid for 1 hour | Prepare a new preview if the quote expires. |
| Compare cap | Up to twice the per-call reservation | Compare runs the reference and the selected pipeline on one input. |
| Evaluation example import | 1–400 examples, 5 MiB | JSON array or JSONL. |
| Trace file import | 10 MiB | Uploaded files expire after 24 hours; copied drafts remain. |
| Task description | 16 KiB | Describe a task tab. |
| Workspace storage | 5 GiB | Drafts, examples, evidence and execution records. |
| Unactivated alternatives | Expire after 30 days | Activated revisions do not expire. |
API keys#
| Limit | Value | Notes |
|---|---|---|
| Key reveal | Once | Copy the key when it is created. Create a new key if it is lost. |
| Per-key spend and rate limits | Optional | Set under Advanced when you create a key. They currently apply to Function calls. |