智能提取处理 (IXP)

Agents can use custom built IXP models for document data extraction. By integrating IXP as a tool, agents can delegate the extraction tasks to state of the art extraction models and use the output for further reasoning or subsequent tool calls. By integrating IXP as a tool, agents can interpret document content, extract high-confidence data points, and use these extracted values in reasoning or subsequent tool calls.

Additionally, you can use the Document Validation Station interface as an escalation mechanism to review and validate extractions before or after the agent’s execution.

先决条件

Before adding IXP as a tool:

• Ensure you have an IXP project and a published model. For details, refer to Creating a project.
• The models must have a deployment tag to be made available to the agent (staging or live). For details, refer to Managing published projects.
• If your IXP model consumes file inputs, confirm your agent is configured to pass or produce attachments as input arguments.

Adding an IXP project as a tool

1. From your agent canvas, open the Toolbox panel.
2. Select IXP.
   
   

   

   
   
3. Choose your IXP project and select the published model.
   
   

   

   
   
4. Specify the model version, identified by tag (staging or live).
5. Provide a tool description that aligns with the intended system and user prompts for the agent.
   
   

   

   
   
6. Open the taxonomy viewer to inspect the fields your model extracts. Use this to guide how your agent should interpret or act on specific field groups or values.

   
   

   

   
   
7. If your agent processes document files:
   • Define an input argument of type File in the Data Manager panel. Alternatively, use another tool or workflow that outputs a file attachment, which the IXP tool can consume.
   • Optionally, add a specific description for the attachment field if the agent handles multiple files during runtime.

Validation Station escalation

You can enable human validation of extracted data through Document Validation Station. Validation Station is an out-of-the-box action app that allows you to review, validate, and modify extractions executed by the model. As with other escalations, any data modified by the reviewer is sent back to the agent as it continues its trajectory.

Note: Decide whether to trigger the validation station mid-trajectory (before other actions) or after the agent completes execution.

To configure:

1. Add an escalation node to your agent.
2. Choose escalation type Document Validation Task.
3. Link it to the specific IXP tool instance.
4. Select a storage bucket for validation results (used for auditing and re-use).
5. Optionally, define a task title and priority.

Testing and debugging

After you're done configuring your agent, run the agent in Debug mode and review traces for the IXP tool execution. The trace output includes the full schema defined in your IXP project, showing input attachments, field groups, and confidence scores.

If you've added a validation escalation, check out the trace metadata. It includes a link to the Action Center task, automatically created for validation.

In Action Center, assign the task to yourself, then check extracted fields, view confidence scores, and manually update values, if needed. When submitted, the updated data is sent back to the agent for continued processing.