universal-ai

Universal Ai

Universal AI endpoint for synchronous non-LLM AI features.

Model format: feature/subfeature/provider[/model]

Supported features:

text/ai_detection, text/moderation, etc.
ocr/ocr, ocr/identity_parser, etc.
image/generation, image/background_removal, etc.

For async features (e.g., audio/speech_to_text_async), use POST /v3/universal-ai/async.

Request body:

model: Model string in format feature/subfeature/provider[/model]
input: Feature-specific input parameters
fallbacks: Optional list of fallback provider strings (max 3); not supported on audio/tts or image/face_recognition, since a voice or a face collection belongs to one provider
provider_params: Optional provider-specific parameters
show_original_response: Include raw provider response (default: false)

Example:

{
    "model": "text/ai_detection/openai/gpt-4",
    "input": {"text": "Analyze this text"}
}

Request body

tagsunknown

modelstring required

Model in format: feature/subfeature/provider[/model]

provider_paramsobject nullable

Native parameters in the provider's own request format, forwarded to the provider alongside input. On some features each provider accepts only an allowlist of parameters, and any other returns 400.

inputobject required

Feature-specific input fields, nested here rather than at the top level of the body. The fields accepted for the feature/subfeature named in model are its input_schema, served by GET /v3/info/{feature}/{subfeature}. Examples:

text/ai_detection: {'text': 'content to analyze'}
ocr/ocr: {'file': '<file ID or URL>', 'language': 'en'}
image/generation: {'text': 'prompt', 'resolution': '1024x1024'}
translation/document_translation: {'file': '<file ID or URL>', 'target_language': 'fr'}

fallbacksstring[]

Fallback providers to try if the primary provider fails, in order. Accepted formats:

'provider' or 'provider/model' (e.g. 'amazon', 'openai/gpt-4')
full model id (e.g 'text/moderation/google', 'image/generation/minimax/image-01')

show_original_responseboolean nullable

Include raw provider response in the output

output_format'full' | 'compact'

compact trims output for LLM and agent callers: base64-encoded media is dropped wherever it appears (the *_resource_url download link next to it stays), and so is dense position data such as per-word OCR boxes, table cell boxes and face landmarks; one box per detected item stays. Fields the output schema marks required may therefore be absent. original_response is not trimmed. full returns the output unchanged. The response echoes the format. Applies to this response only: an async job keeps its full output, so pass output_format=compact again when fetching it.

Example request

{
  "tags": {
    "client": "acme",
    "project": "invoices"
  },
  "model": "text/ai_detection/openai/gpt-4",
  "input": {
    "text": "Analyze this text for AI detection"
  }
}

Response

Successful Response

status'success' | 'fail' required

Whether the request succeeded or failed

coststring required

Cost in credits for this request

providerstring required

Provider name that processed the request

featurestring required

Feature category (e.g., text, ocr, image)

subfeaturestring required

Specific subfeature (e.g., ai_detection, sentiment)

output_format'full' | 'compact'

Whether output, the normalized output, is whole (full) or trimmed for LLM callers (compact); echoes the request parameter. Unrelated to original_response, the raw provider payload behind show_original_response.

outputunknown required

errorobject nullable

Error details from the provider (only present when status is 'fail')

original_responseunknown

edenai_metadataobject nullable

How the request was served: the attempts in order and which one answered. Only present when the request sends x-edenai-metadata: enabled

Universal Ai

Request

Request body

Example request

Response 200OK

Changes

Response