Get traces

Get list of traces with pagination and filtering. See Traces. The REST path /inferenceResults and the inferenceResult* JSON fields keep the old name of this entity.

GET /inferenceResults

Get traces

cURL

curl --request GET \
  --url https://api.galtea.ai/inferenceResults \
  --header 'Authorization: Bearer <token>'

200

[
  {
    "id": "ir_123",
    "sessionId": "session_123",
    "userId": "user_123",
    "index": 0,
    "status": "PENDING",
    "error": null,
    "input": {
      "user_message": "User input text"
    },
    "actualOutput": "Model response",
    "latency": 150,
    "inputTokens": 100,
    "outputTokens": 50,
    "cacheReadInputTokens": 20,
    "tokens": 150,
    "costPerInputToken": 0.00001,
    "costPerOutputToken": 0.00003,
    "costPerCacheReadInputToken": 0.000005,
    "cost": 0.001,
    "creditsUsed": 1,
    "droppedSpanCount": 2,
    "conversationSimulatorVersion": "1.0.0",
    "traceId": "a1b2c3d4e5f6a1b2c3d4e5f6a1b2c3d4",
    "createdAt": "2023-11-07T05:31:56Z",
    "deletedAt": "2023-11-07T05:31:56Z",
    "comments": [
      {
        "id": "comment_123",
        "inferenceResultId": "inferenceResult_123",
        "text": "Agent hallucinated the price.",
        "tags": [
          "hallucination",
          "tone"
        ],
        "userId": "user_123",
        "userName": "John Doe",
        "userEmail": "[e-mail]",
        "createdAt": "2023-11-07T05:31:56Z"
      }
    ]
  }
]

Authorizations

Authorization · string · header · required

API key authorization. Pass your API key in the Authorization header as a Bearer token. Both new (gsk_*) and legacy (gsk-) API keys are accepted, e.g. Authorization: Bearer gsk_... or Authorization: Bearer gsk-....

Query Parameters

ids · string[]

Filter by trace IDs

sessionIds · string[]

Filter by session IDs

evaluationIds · string[]

Filter by evaluation IDs

limit · integer

Maximum number of results

offset · integer

Number of results to skip

fromCreatedAt · string<date-time>

Filter traces created at or after this timestamp (ISO 8601 format)

toCreatedAt · string<date-time>

Filter traces created at or before this timestamp (ISO 8601 format)

sort · string[]

Sort instructions (field and direction pairs). A caller gets 400 for a sort on a field it sees as null on any product the list reaches (evaluation.readInference).

Response

200

application/json

Traces retrieved successfully

id · string

Example:

"ir_123"

sessionId · string

Example:

"session_123"

userId · string | null

Example:

"user_123"

index · integer

Order index within the session

Example:

0

status · enum<string>

Available options: PENDING, GENERATED, FAILED, SKIPPED

Example:

"PENDING"

error · string | null

Error message explaining why this trace failed or was skipped

Example:

null

input · object | null

Structured input data. For plain text input, format is { user_message: "..." }. A content array of parts (audio or file) can sit beside an optional user_message. Each element of a content array is read on its own. A file part is a Galtea file when its uri names a Galtea storage key (files/<organization>/<32 hex>.<extension>, on any host): it is stored signature-free and checked as on a test case input (at most 20 files and 20 MB, an accepted file type, a file of this organization), otherwise 400. A Galtea file missing from storage is stored signature-free without its sizeBytes, and every other element as sent. A metric scores a Galtea file only when its judge model reads that kind of file (PNG, JPEG, WebP, GIF or PDF); any other metric that reads this side is skipped, except a human evaluation or self-hosted one, and the reason names the file.

Example:

{ "user_message": "User input text" }

actualOutput · string object

The agent turn output. Plain text by default (the scored scalar). An envelope { assistant_message, content: [...] } carries a content array of parts (audio or file) beside an optional assistant_message — the scored scalar stays the canonical transcript, and an answer that is only a file carries no scalar at all. A metric scores a Galtea file only when its judge model reads that kind of file (PNG, JPEG, WebP, GIF or PDF); any other metric that reads this side is skipped, except a human evaluation or self-hosted one, and the reason names the file.

Example:

"Model response"

latency · integer | null

Example:

150

inputTokens · integer | null

Example:

100

outputTokens · integer | null

Example:

50

cacheReadInputTokens · integer | null

Example:

20

tokens · integer | null

Total tokens

Example:

150

costPerInputToken · number | null

Example:

0.00001

costPerOutputToken · number | null

Example:

0.00003

costPerCacheReadInputToken · number | null

Example:

0.000005

cost · number | null

Example:

0.001

creditsUsed · integer | null

Example:

1

droppedSpanCount · integer | null

How many spans sent with this trace were refused, for example a span with an empty name or a negative latency. Null means none were refused.

Example:

2

conversationSimulatorVersion · string | null

Example:

"1.0.0"

traceId · string | null

W3C trace ID Galtea sends in the traceparent header on this Direct Inference turn. Stored when the turn starts. This is the OpenTelemetry trace id, not the id of this Galtea trace, which is id.

Example:

"a1b2c3d4e5f6a1b2c3d4e5f6a1b2c3d4"

createdAt · string<date-time>

deletedAt · string<date-time> | null

comments · object[]

Free-form reviewer annotations on this turn, oldest first.

Show child attributes