API docs by Redocly

Evaluator API (v1)

Download OpenAPI specification:Download

Evaluator ™ API for text classification...; receive a decision with detailed policy/constraint signals.

Core

Core decision and warm APIs

Warm the model

Warms the model and dependencies so that the first real predict call is fast. Idempotent.

Authorizations:
subscriptionKey

Responses

Response samples

Content type
application/json
{
  • "ready": true
}

Evaluate text

Runs the Evaluator model on one or more input texts.

Latency and rate limiting:

  • This endpoint is designed for low-latency serving via Azure Front Door.
  • Under early benchmark conditions we observe end-to-end p50 ≈ 230 ms and p95 ≈ 260 ms for typical text-only requests. Actual performance may vary by region, workload, and payload.
  • The service is rate-limited per subscription and tier and may return HTTP 429 (Too Many Requests) when limits are exceeded.

Smoke testing guidance:

  • For a basic connectivity or smoke test, call POST /v1/warm once, then send 10–20 POST /v1/predict requests with a small test payload.
  • Add short delays between calls if you want to avoid triggering rate limits on a single key.

Vector payloads and decision steering:

  • The optional vectors field supports policy-based decision steering with bounded shapes (up to 10 policy dimensions and up to 28 constraint dimensions).
  • Requests that exceed these documented shapes or tier limits are not supported and may be rejected.

Usage tiers:

  • Tier 1 – Decision (Text-only, Trial): send only texts[]. Returns a single decision (YES/NO/TBD) and confidence per text; vectors may be ignored for this tier.
  • Tier 2 – Steered Decision (6 support vectors): send texts[] plus a compact vectors object (policies/constraints and weights). Steering primarily uses the first 6 policy dimensions.
  • Tier 3 – Steered Decision (38 support vectors): send texts[] plus full steering vectors. Up to 10 policy and 28 constraint dimensions may be used for detailed policy/constraint signals in governance use cases.
Authorizations:
subscriptionKey
Request Body schema: application/json
texts
required
Array of strings non-empty

One or more input texts to evaluate.

VectorPayload (object) or object

Optional steering vectors. For Tier 2, a compact 6-vector shape is used. For Tier 3, full 38-dimension vectors can be supplied.

alpha
number <float>
Default: 2

Global steering strength. Higher values increase the influence of vectors over the base model.

v10_min
number or null <float>

Optional minimum threshold for policy (v10) activation.

e28_min
number or null <float>

Optional minimum threshold for constraint (e28) activation.

text_conf_max
number <float>
Default: 0.8

Maximum base model confidence before steering adjustments are limited.

include_probs
boolean
Default: true

If true, include per-class probability arrays in the response.

include_explain
boolean
Default: true

Reserved for future detailed explanation payloads.

max_length
integer <int32>
Default: 128

Maximum tokenized length per input text.

Responses

Request samples

Content type
application/json
{
  • "texts": [
    ]
}

Response samples

Content type
application/json
{
  • "decision": [
    ]
}

Health

Health and readiness checks

Private health check

Liveness/readiness check for the Evaluator API process.

Responses