generated: '2026-07-19'
method: searched
source: https://docs.nvidia.com/sdgm/rest-api , https://docs.nvidia.com/sdgm/rfm/configuration
  , https://docs.nvidia.com/sdgm/quotas-and-limits
authentication:
  style: api-key-header
  header: X-API-Key
  key_format: <customer_id>:<secret_value>
  artifact: authentication/kumoai-authentication.yml
base_url:
  pattern: https://<customer_id>.kumoai.cloud/api
  note: Base URL is tenant-scoped; there is no shared multi-tenant public host.
query_interface:
  name: Predictive Query Language (PQL)
  description: SQL-like declarative language for expressing a prediction as Target
    × Entity × Horizon (e.g. PREDICT SUM(orders.amount, 0, 30, days) FOR EACH users.user_id).
    Predictions are requested as queries rather than as per-resource CRUD calls.
  docs: https://docs.nvidia.com/sdgm/rfm/querying-rfm
idempotency:
  supported: false
  note: Kumo publishes no idempotency-key contract for the REST API. Retrain and batch-prediction
    calls are job-submitting operations without a documented request-deduplication
    header, so no Idempotency pointer is emitted.
pagination:
  documented: false
  note: The REST API surface is job-oriented (retrain, batch predict, job status, cancel)
    rather than collection-oriented; no pagination contract is published.
async_jobs:
  model: submit-and-poll
  operations:
  - retrain a predictive query
  - run a batch prediction
  - monitor job status and evaluation metrics
  - cancel an in-progress job
  queueing: Up to 15 jobs (training or batch prediction) may be queued.
  parallelism: Up to 10 parallel jobs on SaaS/Databricks Native; 1 on Snowflake Native.
  docs: https://docs.nvidia.com/sdgm/rest-api
retries:
  documented: true
  note: The RFM SDK exposes retry handling and batch mode as runtime configuration.
  docs: https://docs.nvidia.com/sdgm/rfm/configuration
run_modes:
  parameter: run_mode
  values:
  - id: DEBUG
    context_size: 100
    use_case: Quick iteration, testing queries
  - id: FAST
    context_size: 1000
    default: true
    use_case: Good balance of speed and quality
  - id: NORMAL
    context_size: 5000
    use_case: Higher quality predictions
  - id: BEST
    context_size: 10000
    use_case: Maximum quality
  docs: https://docs.nvidia.com/sdgm/rfm/configuration
temporal_semantics:
  anchor_time: Prediction timestamp; defaults to the maximum timestamp in the data,
    or per-entity when set to "entity".
  context_anchor_time: Maximum anchor timestamp for in-context examples.
  use_prediction_time: Include the anchor time as an additional feature.
  lag_timesteps: Number of past timesteps included as lagged target features.
  note: Anchor time is the core correctness control for preventing data leakage.
  docs: https://docs.nvidia.com/sdgm/troubleshooting/what-is-anchor-time-and-why-is-it-important
errors:
  envelope: PQL parser/validator error messages plus structured failure categories
    (A–F).
  artifact: errors/kumoai-problem-types.yml
  docs: https://github.com/kumo-ai/kumo-coding-agent/blob/main/context/platform/pql-errors.md
rate_limits:
  style: quotas-and-concurrency (data scale limits + job queue/parallelism), not request-rate
    throttling
  artifact: rate-limits/kumoai-rate-limits.yml
  docs: https://docs.nvidia.com/sdgm/quotas-and-limits
versioning:
  scheme: semver on the Python SDK and MCP server packages; no version segment in
    the REST API path
  artifact: lifecycle/kumoai-lifecycle.yml
cross_references:
  authentication: authentication/kumoai-authentication.yml
  errors: errors/kumoai-problem-types.yml
  lifecycle: lifecycle/kumoai-lifecycle.yml
  rate_limits: rate-limits/kumoai-rate-limits.yml