LLMs in ABBYY Vantage

ABBYY Vantage uses large language models (LLMs) to extract, classify, and understand document data from natural-language instructions. There are two ways to reach a model:

• ABBYY Phoenix Plus: an LLM endpoint that ABBYY supplies and operates for you. Nothing to obtain, configure, or rotate.
• Your own LLM connection: credentials for your own account with an external provider such as Anthropic, Google AI Studio (Gemini), Microsoft Foundry, Mistral AI, or OpenAI. Custom (OpenAI-Compatible) covers an OpenAI-compatible endpoint you host or control yourself.

Both are tenant connections of the same kind. Both appear under ADMIN → Configuration → Connections, and both can be selected by a prompt-based activity in a Document Skill. A tenant can use both at the same time.

​

Choosing a source

​

ABBYY Phoenix Plus

ABBYY Phoenix Plus gives a tenant LLM capability without a provider account of its own. There is no API key to obtain, no endpoint URL or model name to enter, and no credential to rotate. ABBYY selects, hosts, and maintains the model, and can update it without any change on your side. When the entitlement is enabled, a connection named ABBYY Phoenix Model appears under ADMIN → Configuration → Connections, shown with a padlock icon and the description Managed by ABBYY. Because ABBYY owns the credential behind it, the connection is read-only: you can test it, but you cannot edit or delete it, and its credentials are never displayed, not even masked.

​

Two ways to use it

The connection is one thing. What you can do with it depends on where you use it. If your goal is LLM extraction with the least work, use the prompt-based activity and select ABBYY Phoenix Model from the LLM Connection dropdown. For the steps, see Extract data with prompt-based activities. A shared connection is not a shared capability. The techniques in Call ABBYY Phoenix Plus from a Custom activity are script work. Sending OCR JSON or page images, validating the response, and taking field positions from the OCR layer have no equivalent in the prompt-based activity. That activity builds the prompt and maps the response for you.

​

Calling it from a script

Phoenix Plus can also be invoked from a Custom activity inside a Process skill, so the model runs as a step in your workflow with the payload and the response under your control. No HTTP plumbing, authentication, or provider-specific integration code is required. That work is what the entitlement removes. You still write the activity script itself, and two constraints shape how you write it:

• Each message has a size limit that counts image content, so large page images must be managed rather than assumed.
• The model reads and classifies. It does not measure. Take field positions and bounding boxes from the Vantage OCR layer, never from the model.

Both are covered in Call ABBYY Phoenix Plus from a Custom activity.

​

Using your own LLM connection

A tenant administrator adds credentials for a provider account you hold, and skill designers select that connection in a prompt-based activity.

​

When to use LLMs

Prompt-based LLM extraction needs no training data and adapts quickly to new fields and layouts. Each extraction calls a model, so it costs more and runs slower than a trained model, and results can vary slightly between runs.

​

Data handling and security

When a skill uses your own LLM connection, the document content is sent to the external provider you connect to and is processed on that provider’s infrastructure, subject to the provider’s terms and data policies. Choose a provider that meets your data-residency and compliance needs. Microsoft Foundry supports Azure-based data residency and compliance controls. With Custom (OpenAI-Compatible), content goes to the endpoint you name in the connection, so data residency follows wherever you run that endpoint. When a skill or script uses ABBYY Phoenix Plus, document content is sent to the model endpoint ABBYY operates for this service rather than to a provider account of your own.

​

Where Phoenix Plus processes your data

Phoenix Plus runs in three regions, and your tenant uses the region it is deployed in. You do not choose the region per request. In the United States and the European Union, requests are processed within that geography, pinned to specific data centres inside it.

​

How your content is used

Content you send to ABBYY Phoenix Plus, and the responses it returns:

• Is not available to other customers.
• Is not available to the provider of the underlying model.
• Is not used by that provider to improve their models.
• Is not used to train any generative AI foundation model without your permission.
• Is not used to improve any first-party or third-party product without your explicit permission or instruction.

​

What is retained

ABBYY does not retain the content you send to Phoenix Plus, or the responses it returns. The service is designed not to hold them. The regional table above describes where processing happens and where any storage at the platform layer would reside. It does not mean ABBYY keeps a copy of your documents. What your own skills do with the data is separate and under your control. A Custom activity script that writes a model response into a document field, or logs it, stores that content in your tenant in the ordinary way, subject to your own data-retention settings. Protect the credentials stored in each connection you configure yourself:

• Never share API keys or include them in a connection’s description.
• Rotate credentials regularly.
• Use separate connections for development and production.
• Grant each key the least privilege it needs.

These steps do not apply to the ABBYY Phoenix Model connection, which stores no credential of yours and never displays the credential it uses.

​

Cost and token usage

Usage of your own LLM connection is billed by the provider, based on the number of tokens processed. Cost scales with document length, prompt complexity, the document format sent to the model, and the number of fields you extract. Monitor usage in your provider’s dashboard, and use smaller, faster models for simple tasks. ABBYY Phoenix Plus is not billed by an external provider. Access is a contracted entitlement arranged with your ABBYY account team, and consumption is metered by ABBYY. Your own LLM connections are not metered this way, because they bill through your provider instead. Phoenix Plus consumption is not currently shown in the Vantage interface; for consumption figures, contact your account team. Because usage draws on your entitlement rather than a provider account you control, consider the volume before pointing a bulk reprocessing job at a skill that uses it.

​

Supported providers and models

Vantage supports several LLM providers for connections you configure yourself, each with its own connection fields and models. For the current list, see Supported LLM providers and models. ABBYY Phoenix Plus does not appear there, because it has no connection fields to configure and its model is selected by ABBYY.

​

What’s next

​

Related topics

• Skill Designer overview
• Document skill parameters