OCR through text models

Besides the dedicated OCR Model entity, OCR can also be performed through a regular LLM (text) provider. AlphaEdge 🇫🇷 🇪🇺 can be configured as a standard LLM provider: you call the OpenAI-compatible /chat/completions endpoint with a message that contains an image or PDF content part, and the assistant response is the extracted text.

This is convenient when you already have a chat/LLM pipeline and want OCR to flow through the exact same surface: the standard /chat/completions endpoint, your existing OpenAI clients/SDKs, model constraints, prompt contexts, caching, budgets, and observability.

Text model vs OCR Model

	OCR through text models	Dedicated OCR Model
Entity	AiProvider (LLM provider)	OcrModel
Endpoint	POST /chat/completions	POST /ocr
Input	A chat message with an image/pdf content part	JSON document or multipart upload
Output	A standard chat completion (assistant text)	{ pages[].markdown, usage_info, ... }
Best for	Reusing an existing chat pipeline / OpenAI clients	A purpose-built OCR API and the ocr_call workflow function

Both approaches use the same AlphaEdge backend; pick the one that best fits how you consume the result.

Provider configuration

Create an LLM provider with provider set to alphaedge. Authentication uses the X-API-Key header (set through connection.token).

{
  "id": "provider_xxxxxxxxx",
  "name": "AlphaEdge OCR",
  "description": "AlphaEdge OCR as a text model",
  "provider": "alphaedge",
  "connection": {
    "base_url": "https://api-endpoints.alphaedge-ai.com",
    "token": "${vault://local/ALPHAEDGE_API_KEY}",
    "timeout": 180000
  },
  "options": {
    "model": "alpha-digit-max",
    "allow_config_override": true
  }
}

Field	Type	Default	Description
connection.base_url	string	https://api-endpoints.alphaedge-ai.com	AlphaEdge base URL
connection.token	string	—	AlphaEdge API key (sent as X-API-Key). Supports vault references and comma-separated rotation.
connection.timeout	number	180000	Request timeout in milliseconds
options.model	string	alpha-digit-max	The default OCR model (alpha-digit-max or alpha-digit-medium)
options.pdf_password	string	—	Optional password for protected PDFs
options.allow_config_override	boolean	true	Allow the request body to override options (e.g. model, pdf_password)

Expose the provider on a route with an LLM proxy (the OpenAI Compatible API plugin or the chat completions proxy). See Expose a provider.

Usage

Send a chat completion whose message contains a file content part. A file (image or PDF) content part is required — a request without one returns 400.

Image input

curl https://my-llm-endpoint.example.com/chat/completions \
  -H "Content-Type: application/json" \
  -H "Authorization: Bearer $OTOROSHI_API_KEY" \
  -d '{
    "model": "alpha-digit-max",
    "messages": [
      {
        "role": "user",
        "content": [
          { "type": "image_url", "image_url": { "url": "data:image/png;base64,iVBORw0KGgo..." } }
        ]
      }
    ]
  }'

The image can also be passed as a remote https:// URL.

PDF input

PDFs are passed as a document content part:

{
  "model": "alpha-digit-max",
  "messages": [
    {
      "role": "user",
      "content": [
        {
          "type": "document",
          "source": {
            "type": "base64",
            "media_type": "application/pdf",
            "data": "JVBERi0xLjcK..."
          }
        }
      ]
    }
  ]
}

Response

The result is a standard chat completion. The extracted text is the assistant message content:

{
  "id": "chatcmpl-xxxxx",
  "object": "chat.completion",
  "model": "alpha-digit-max",
  "choices": [
    {
      "index": 0,
      "message": { "role": "assistant", "content": "The full extracted text..." },
      "finish_reason": "stop"
    }
  ],
  "usage": { "prompt_tokens": 0, "completion_tokens": 0, "total_tokens": 0 }
}

Notes

• Required file part: the call fails with 400 if no image or PDF content part is present in the messages.
• Streaming: AlphaEdge OCR is not natively streamed. When stream: true is requested, the single extracted text is returned as one streamed chunk.
• PDF password: pass pdf_password in options, or in the request body when allow_config_override is enabled.
• Token usage: OCR does not consume LLM tokens, so usage counters are reported as 0.

See also

• OCR Models — the dedicated OcrModel entity and POST /ocr API
• Speech-to-Text — AlphaEdge also supports audio transcription