- Overview
- Model building
- Model validation
- Model deployment
- Frequently asked questions

Unstructured and complex documents user guide
You can configure the underlying LLM as well as its settings in the Model configuration option from the Build tab.
- Intelligent pre-processing:
- None
- Table model - mini
- Table model
- Extraction model:
- GPT-4o
- Gemini
- Advanced options:
- Attribution
- Temperature
- Top P
- Seed
- Frequency penalty
- Prompt override
Adjust these settings to improve the accuracy of the model's predictions and enhance its performance.
Intelligent pre-processing options improve prediction performance when documents are difficult for models to interpret, due to complex formatting.
- None - This default option is suitable for most documents that do not have tabular content.
- Table model - mini - Optimized for tabular content and latency. This option is best suited for documents with simple tables or multiple tables.
- Table model - Optimized
for more complex tabular content. This option is best suited for documents with
complex nested tables, tables with merged cells, bullet points, or tables
spanning across multiple pages.
Note:
- While this performs best on complex tables, it increases the latency of predictions.
- This feature relies on Gemini models through the AI Trust Layer.
Example of Intelligent pre-processing
this
period
are confused for the ones in the year to date
column.this period
and year
to date
, are extracted correctly.The Extraction model option represents the underlying LLM used for extraction.
- GPT-4o
- Gemini
Choosing the most suitable model
Different models will perform differently for different use cases, but you are recommended to use Gemini where possible. Several other pre- and post-processing features, which help optimize performance and user experience, are also Gemini-based.
GPT-4o has a restriction of 50 pages and can only process more using the currently previewed iterative calling feature. Gemini does not have the same restriction and can process documents in IXP up to 200 pages in a single call. The Gemini limit may vary slightly based on the density of field values within the document.
In addition, the Gemini model has an input limit of 200 pages by default, compared to the 50 pages input limit of GPT-4o.
Switching from one model to another
To switch from one model to another, use the dropdown list of the Extraction model option and select Save. This will trigger a new project version to be created and new predictions to be generated automatically.
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 the performance metrics in Measure.
Advanced options allow you to customize the settings for your models, select which attribution method to use, and use the prompt override.
Expand the setting to view all available options:
- Attribution - The method used
for attributing predictions to the relevant part or text in the document. Select one
of the following options:
- Rules-based - Uses an extensive set of rules and heuristics to match the correct spans on a page to the predicted values from the model. This is a low-latency option, but it sacrifices performance in terms of successful attributions compared to the model-based option.
- Model-based - Uses an additional LLM call to successfully match the predicted values to the correct spans on the page, as these values can often be repeated in different parts of the page. This is the most performant option in terms of successful attributions, but it does add some latency to predictions. This option relies on using Gemini models.
- Temperature - The sampling temperature to use. Select a number between 0.0 and 2.0. Higher values make the output more random.
- Top P - Samples only from
tokens with the
top_p
probability mass. Select a number between 0.0 and 1.0. - Seed - If specified, repeated requests with the same seed and parameters should return the same result.
- Frequency penalty - Select a number between -2.0 and 2.0. Positive values reduce the probability of the model repeating tokens that have already appeared in the text.
- Prompt override - Overrides the default system prompt with a new value. This option is disabled by default. Once enabled, the Append task instructions prompt and the Append field instructions prompt options are enabled for configuration.