- Overview
- Introduction
- Extracting data from unstructured documents
- Building and deploying models
- Quotas
- Auditing LLM interactions
- Model building
- Model validation
- Model deployment
- Consuming models
- API
- Frequently asked questions
AI Trust Layer audit for Unstructured and Complex Documents; which LLM inputs and outputs are recorded for document extraction operations, and how to review them for compliance.
The Unstructured and Complex Documents capability sends all its Large Language Model (LLM) requests through the UiPath AI Trust Layer™, which records each document extraction operation in an audit log. The audit log gives administrators visibility into the data sent to LLMs and the responses they returned, supporting compliance, governance, and data transparency requirements.
View audit entries
Each extraction appears in the Audit tab of the AI Trust Layer page, under the IXP product name, including the inputs sent to the Large Language Model (LLM) and the outputs it returned, for both design-time and run-time requests.
Entries remain visible for 60 days, and you can export them for longer retention periods. For details about the audit view, exports, and retention, refer to Viewing audit logs.
How extraction operations are recorded
Each document processing job is recorded as a single audited action. A job can involve multiple underlying LLM requests, such as pre-processing followed by extraction, but these are displayed together as one entry.
Extraction operations are audited in both contexts where predictions are generated:
- Design-time: predictions generated while building and validating the model, such as in the annotation view.
- Runtime: predictions requested by published models, through Studio workflows or the API.
Audited inputs
The input recorded for an extraction operation includes the following:
- Project ID.
- Model version.
- File name of the processed document.
- Request type: design-time or runtime.
- Prompt, built from the taxonomy and field instructions.
- Model configuration: the pre-processing option and the advanced options (attribution method, temperature, Top P, seed, and frequency penalty).
Audited outputs
The output recorded for an extraction operation is the raw LLM response for the extraction request, containing the predictions returned for the document.
What is not audited
Audit entries do not include the following:
- Input and output for pre-processing or post-processing steps.
- Implementation details on value-adding features or differentiators. For example, intelligent pre-processing options, attribution implementation details.
Control audited data
Automation Ops policies for the AI Trust Layer control how audit data is handled: you can turn off the saving of inputs and outputs, or mask personally identifiable information (PII) before requests reach the LLM. For details, refer to the following resources: