This feature is in Limited Availability. Existing customers who are interested in accessing this feature should reach out to our support team for assistance – support@rossum.ai
Reasoning fields are used to extract specific values from other fields’ content.
Overview
Rossum’s standard AI engines can extract values exactly as they appear in documents. Formula fields allow you to manipulate those values using predefined rules. Reasoning fields go a step further: it enables the AI to interpret the document context and generate new values based on your instructions.
With reasoning fields, you can define what the AI should infer and what information it may use to do so. Our AI can either generate a new value from context or select one from a predefined list.
Importantly, reasoning fields can learn from your manual corrections within a queue, helping the AI improve its accuracy over time: If it makes a mistake once, it won’t repeat it.
How to create a Reasoning Field
You can add a reasoning field in the Field Settings of a queue by selecting a Reasoning Value Source.
A dedicated section allows you to define the reasoning context, the data available to the AI for making inferences. This includes:
Relevant fields: Input fields that the reasoning field will consider for its answer.
Relevant properties (optional): Additional information (Label, Description) that can give reasoning more context about your task.
Instructions (optional): Label and relevant fields are often enough. If the output isn’t accurate, you can provide a more detailed task description.
How to write instructions?
Be specific, precise, and unambiguous.
Always mention field_ids in your instructions.
Say what to ignore or exclude
Include clear examples to illustrate the expected output. (1-3 are often enough.)
Define the exact output format, provide a fallback.
Ask another AI assistant to suggest improvements to your instructions.
Example
Bad instructions:
Return the country code, based on the address.
Good instructions:
Return the ISO 3166-1 alpha-3 country code (e.g., CZE, AUT, NOR) for the given address.
Output: 3 uppercase letters only
If the country is missing, infer from context (city, postal code, language, etc.)
If unclear, return "UNK"
Example: "KÅ™ižÃkova 148/34, 186 00 KarlÃn" → CZE
Testing your setup
Before saving, test your reasoning field using the built-in testing tools:
View real documents from your queue in the testing section.
Use the preview row to test with custom values.
Click Update values after making changes to refresh the test results.
Common use cases
Extract early payment discounts. Often, the early payment discount can come in various shapes and languages. Reasoning gets you covered!
Convert unstructured line items into a structured format. (For example, calculate total amounts from complex item descriptions)
Extract header fields from email bodies. (Create a new formula field. Ask it to “extract email body“. Then feed it into reasoning.)
Analyze email sentiment. (Identify dissatisfaction or urgency in document delivery emails.)
Categorize the goods in line items.
Reasoning vs Formula Fields
Use formula fields when you can clearly express the transformation logic. (For example, remove all spaces from the IBAN field or calculate sum from all items in a table.)
Use reasoning fields when the task involves ambiguity, variability, or contextual interpretation.
Note: Formula fields are generally faster and more cost-efficient.
Requirements
The user must have access to Queue Settings. (typically available to administrators or managers).
Formula fields must be enabled.
Reasoning fields must be enabled.