Evaluations

Run a typed evaluation

Jev bewertet einen JSON-Zustand anhand benannter Fragen und gibt typisierte Antworten als boolean, choice oder score zurück. Rufen Sie POST /v1/evaluations mit Bearer-Authentifizierung auf; das Kontextfenster umfasst 32K Tokens. Jev finden Sie über /v1/models?output_modalities=decisions oder /v1/models?output_modalities=all. Die standardmäßige reine Text-Modellliste enthält keine Modelle, die ausschließlich decisions ausgeben. Jev ist nur über den evaluations-Endpunkt verfügbar; chatförmige Endpunkte antworten mit HTTP 400. Der aktuelle Listenpreis beträgt $0.042 je 1 Mio. Eingabe-tokens und $0 je 1 Mio. Ausgabe-tokens. Es müssen ausreichend Credits verfügbar sein.

POST/v1/evaluations

Autorisierung

Authorizationerforderlich

string · header

API-Schlüssel als Bearer-Token im Authorization-Header.

Body

model

string

The optional evaluation model in the request body. Omit it to use the default typesafe-ai/jev; any other value is rejected as model_not_available.

optional; defaults to typesafe-ai/jev

Example: "typesafe-ai/jev"

stateerforderlich

JSON string | object | array

The state or evidence to evaluate. It may be a string, object, or array.

JSON value

Example: "The support agent issued a full refund."

questionserforderlich

record<string, Question>

One or more named questions evaluated in parallel.

typeerforderlich

boolean | choice | score

The typed answer requested for this question.

instructionserforderlich

JSON string | object | array

Instructions describing what this question should decide.

criteria

record<string, JSON> | JSON[]

Required for choice questions as an option-to-description record; required for score questions as an ordered array of at least two criteria; optional for boolean questions.

POST /v1/evaluations

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
curl https://api.bazaarlink.ai/v1/evaluations \
  -H "Authorization: Bearer $BAZAARLINK_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "typesafe-ai/jev",
    "state": "The support agent issued a full refund and apologised for the delay.",
    "questions": {
      "refunded": {
        "type": "boolean",
        "instructions": "Was a refund issued?"
      },
      "tone": {
        "type": "choice",
        "instructions": "Tone of the agent?",
        "criteria": {
          "apologetic": "The agent apologises",
          "neutral": "No emotional language",
          "hostile": "The agent is rude"
        }
      },
      "quality": {
        "type": "score",
        "instructions": "How complete was the resolution?",
        "criteria": ["poor", "partial", "complete"]
      }
    }
  }'

Antwortbeispiele

Success. The response includes id, object, model, answers, usage, rounding, and warnings. answers is keyed by question name; choice probabilities use option names as keys. For score questions, score is the expected value over indices 0 through N-1, and probabilities keys are the string indices in criteria order. usage reports input_tokens, output_tokens, and total_tokens; rounding reports probabilityDecimals and scoreDecimals.

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
{
  "id": "eval-a3d350e5-2705-4a9b-be7c-838fa4e4eb2b",
  "object": "evaluation",
  "model": "typesafe-ai/jev",
  "answers": {
    "refunded": {
      "type": "boolean",
      "probability": 0.99
    },
    "tone": {
      "type": "choice",
      "choice": "apologetic",
      "probabilities": {
        "apologetic": 1,
        "neutral": 0,
        "hostile": 0
      }
    },
    "quality": {
      "type": "score",
      "score": 1.98,
      "probabilities": {
        "0": 0,
        "1": 0.02,
        "2": 0.98
      }
    }
  },
  "usage": {
    "input_tokens": 394,
    "output_tokens": 74,
    "total_tokens": 468
  },
  "rounding": {
    "probabilityDecimals": 2,
    "scoreDecimals": 2
  },
  "warnings": []
}