Configure the extraction strategy, extraction model, language support, layout parsing, attribution method, and overrides of a Flow Extract node.

The model configuration of an Extract node defines how the node processes documents and which LLM it uses to extract data.

You can configure the underlying LLM as well as its settings in the Model configuration section of the Definition tab as follows:

Result: The Extract node uses the new model configuration.

Different models perform differently for different use cases, but Gemini is recommended where possible.

GPT-4o has a restriction of 50 pages.

Gemini 2.5 Flash can process documents in IXP up to 500 pages in a single call, with higher page counts supported in preview. The Gemini limit may vary slightly based on the density of the document.

In terms of field values, Gemini 2.5 Flash has a higher output context window, which allows it to handle more field values than GPT-4o. There is no fixed limit on the number of field values per document. The limit depends on the model, and it may vary depending on the complexity of the document.

To switch from one model to another, select a model from the Extraction model dropdown list, and then select Save.

This creates a new node version. To get predictions from the new model, select Refresh predictions in the annotation view of each document. You can also refresh predictions for multiple documents at once, or run an evaluation on all or multiple documents.

If you need to switch the model for performance reasons, check first whether the alternative model can solve the core problem that the current model cannot solve. If it can, optimize the new model to improve its evaluation metrics.

There is no fixed limit on the number of field values that IXP can extract per document. The limit depends on the extraction model.

If a document requires more field values than the model can return, the model may reach its token limit. This can happen even if the document is well within the page limits listed in this section. If you reach this limit, set the Extraction strategy to Advanced, if it is available to you, to handle more field values.

If token limits are exceeded, predictions may be incomplete or empty regardless of the number of pages. For example, an 80-page document with 1,400 extracted field values may trigger a token-limit warning in the user interface. At runtime, when the token limit is exceeded, the API response will show this by returning occurrence and extraction confidence values of 0.

If page limits are exceeded, only the pages within the specified limits below are processed. Both of these behaviors apply to both Gemini and GPT-4o.

Runtime limits include:

The Language support setting defines which alphabets the OCR is optimized to recognize before extraction. If your documents use any non-Latin scripts, select Extended languages.

The available options are:

You can upload documents and build models in any language that is well-supported by the selected extraction model. The set of supported languages depends on the model that you select.

Extraction accuracy can vary depending on the language of your documents. Results may be less consistent for languages such as:

For these languages, you should validate model performance on a representative sample of your documents before moving to production.

The Layout parsing setting is an optional step that reconstructs document layouts and passes them to the model, to improve extraction performance on documents with tables or complex layouts. This step adds latency to each document.

The available options are:

The Attribution setting defines the method for attributing model predictions to their references in the document.

The available options are:

The Overrides setting enforces overrides to the model configuration, to apply advanced developer settings. Enter the overrides in JSON format in the JSON editor. The editor validates the JSON and shows a Valid label when the content is valid.