Define Extract fields using a constrained JSON Schema, and configure extraction settings
The extraction configuration defines "which information the Agent needs, and how to interpret that information".
Log in to ParseX, go to Field Extraction on the left, and find the field configuration area. Prepare the field structure to extract, also called the schema, which determines which business fields Extract returns, along with their field types and nested structure. A clear, stable schema reduces field ambiguity and makes it easier for downstream systems to consume the results directly. You can think of it as a form to be filled in: you define the columns and filling requirements, and ParseX obtains the corresponding values from the materials.
For the rules and limits of schemas, see the Schema rules reference.
Fill in the field name, type, and description row by row in the field list. To express composite information, use child fields to organize the structure, such as placing the company name and address under the same company information object.
For example, "revenue" can easily correspond to multiple values; "the operating revenue for the reporting period, keeping the original amount and unit, not using the prior-year figure" makes the business scope easier to pin down.
If you have not prepared a field structure yet, you can describe the information you need and let the system generate a candidate schema. The input supports up to 4,000 Unicode characters, which you can generally treat as a per-page character limit.
When you already have field definitions, or need to adjust multiple fields at once, you can enter batch editing to view and edit the complete JSON Schema. After editing, apply it as prompted on the page, then check whether the field list matches your expectations.
The following simple example defines two text fields to illustrate the relationship between "field name, type, and description":
This is an example of a field structure, not an extraction result. When using batch editing, pay attention to quotation marks, commas, and brackets; if the page reports a structure error, fix it before applying. For the full range of support and API parameters, see the standalone API reference.
Generating fields from natural language answers "which fields are needed", while a custom prompt supplements "what requirements this extraction run follows". The meaning of each field should be written in its field description, and unified requirements should be written in the custom prompt; avoid contradictions between the two.
After completing the configuration, submit the extraction task. See Get field extraction results.
Configure the extraction field schema
Log in to ParseX, go to Field Extraction on the left, and find the field configuration area. Prepare the field structure to extract, also called the schema, which determines which business fields Extract returns, along with their field types and nested structure. A clear, stable schema reduces field ambiguity and makes it easier for downstream systems to consume the results directly. You can think of it as a form to be filled in: you define the columns and filling requirements, and ParseX obtains the corresponding values from the materials.
For the rules and limits of schemas, see the Schema rules reference.
Option 1: Set fields one by one
Fill in the field name, type, and description row by row in the field list. To express composite information, use child fields to organize the structure, such as placing the company name and address under the same company information object.
| Setting | Purpose | Tips |
|---|---|---|
| Field name | Identifies an item of information in the result | Use names with clear meanings that are easy to distinguish; avoid giving multiple fields similar names that cannot be told apart |
| Type | Describes the expected value or structure | Choose based on the business purpose; if an amount needs to keep its original unit, validate it as a text type first |
| Description | Specifies the content and scope to look for | Clearly state the subject, period, unit, and concepts that are easy to confuse |
| Child fields | Groups related information together | Only add levels when you need to express a composite structure; avoid meaningless nesting |
Option 2: Generate fields from natural language
If you have not prepared a field structure yet, you can describe the information you need and let the system generate a candidate schema. The input supports up to 4,000 Unicode characters, which you can generally treat as a per-page character limit.
- Open the natural language generation entry.
- Describe the document type, the information needed, and any necessary field requirements.
- Generate the field structure, then review the field names, types, descriptions, and hierarchy.
- Adopt it after confirming it meets your needs, then check the field configuration in the workspace.
- Run the extraction and use real results to verify whether this set of fields is suitable.
Extract the reporting period, operating revenue, and budget commitment utilization rate from operating reports. Keep the original amount and unit for operating revenue, keep the percentage notation for the budget commitment utilization rate, and only take data for the current reporting period.The generated structure does not automatically obtain field values; you get the field values from the materials only after completing the configuration and running the extraction. If a business field is missing, add it to the configuration; if a field's meaning is inaccurate, revise its description first, then run the document extraction.
Option 3: Edit the field structure in batch
When you already have field definitions, or need to adjust multiple fields at once, you can enter batch editing to view and edit the complete JSON Schema. After editing, apply it as prompted on the page, then check whether the field list matches your expectations.
The following simple example defines two text fields to illustrate the relationship between "field name, type, and description":
Configure inference, source evidence, and custom prompts
| Setting | Default state | Usage notes |
|---|---|---|
| Field inference | Off | Enable it only when fields need to be inferred from the document context; distinguish inferred results from information explicitly stated in the source and review them |
| Source evidence | On | Keep it on when you need to trace back to the source; when results include source page numbers, use them to compare against the original text |
| Custom prompt | Use as needed | Adds unified requirements for the run, such as focusing only on a specific business entity or statistical scope |