Session 6 Specialized AI Associate Series: The GenAI Experience in UiPath Document Understanding

🚀 Welcome to Session 6/ AI Associate Developer Series 2025!

In this session, we will discover in depth the latest UiPath Document Understanding features and updates.

📕 Agenda:

Active Learning (Modern Experience, DocPath LLM)
Active Learning Best Practices, comparing with Classic Experience
Consuming Modern Experience Projects (Studio Web or Studio (Preview))
Generative Classifier and Extractor
📥 Practice project samples download: DU project with GenAI

For links to the self-study modules, go here: https://academy.uipath.com/learning-plans/specialized-ai-associate-training, and click on the second tab.

1. Overview of the solution and how it helps with model training Active Learning (Modern Experience)

2. Centric Consulting 6x UiPath MVP Purdue University 1x UIPath MVP Tracy Dixon Will Oprisko Your Speakers

3. Agenda Slide Introduction Active Learning (Modern Experience) Overview of Active Learning Active Learning Best Practices, comparing with Classic Experience Build - Labelling Data and training extraction models Measure and Deploy – Assessing model health and consuming projects Model training and serving for Modern projects Overview of DocPath LLM - how it impacts our solutioning Consuming Modern Experience Projects Studio Web (via process template, data service, integration service) Studio (via UiPath.IntelligentOCR.Activities package, data service) Generative Classifier and Extractor When to use How to use

4. Introduction

5. What does this training cover in the exam curriculum? UiPath Document Understanding • Define DU and identify structured vs unstructured documents • Rule-based vs GenAI models • OCR vs Document Understanding • Classic vs Modern experiences Document Understanding Framework • Build and measure projects using the DU Framework • Use the Process Template and Validation Station Document Understanding Activities • Taxonomy Manager, Digitize, Classify, Extract, Validate • Configure ML and Generative extractors AI Center & Model Building • Active Learning (Modern Experience) • DocPath LLM overview and benefits • Model training and deployment best practices Autopilot • Use GenAI to generate workflows and enhance DU automation

6. Active Learning (Modern Experience) Overview of the solution and how it helps with model training

7. What is Active Learning? Active Learning is a next-gen AI-powered model training experience within UiPath Document Understanding. It involves an iterative process between annotators and a model to minimize the amount of data needed to train a machine learning model. All parts of the training process happens synchronously. The ML algorithm actively selects the most informative samples to label from a pool of unlabeled data (e.g. documents where it’s unsure about extracting certain fields).

8. The goal of Active Learning is to decrease the number of samples required to build models while reducing the time and effort needed for both annotating and training. Active Learning makes it easy for users to track model’s performance and ensure desired business outcomes. Anyone can train AI models—no coding or ML skills required 80% faster model training—from a week, down to just a day Guidance on model optimization—humans & AI collaborating together Instant model evaluation—built-in model performance analytics How does Active Learning help with model training?

9. How is the modern experience different from the classic experience? Build Load samples, annotate, and train your model in a guided experience Upload & classify your documents automatically Use Generative AI to automatically pre - annotate fields for users to validate Follow recommendations that guide you on how to build & improve your model efficiently Model training occurs in the background to see the impact of user actions Measure Understand your model performance and how to improve it Evaluate your project readiness with model metrics Model metrics indicate which parts of the project need most attention Users are guided to specific parts of the project to improve model’s performance using recommended actions See distribution of dataset Publish Deploy and manage your projects with ease Choose and deploy the model in just a couple of clicks Easily version your entire project Start consuming the project you built in Studio, Studio Web, or via APIs Monitor Monitor and audit the performance of your automation Insights dashboard shows the performance of your project in automations Easily drill down into the status and details of documents being processed Visibility into human in the loop events – see who and what has been corrected Search through the history of processed documents

10. How do I access modern projects?

11. Active Learning Best Practices Labelling data and training extraction models

12. Uploading data and automatic classification Once you create a new project, you will be prompted to upload some sample documents. Tip : Always prefix your samples with the vendor/layout name and document type (if you are planning to work with multiple document types in the same project)  If your documents are already sorted into multiple document types: Create each document type first and upload your documents there For each document type, upload at least 30 samples If your documents are not sorted : Drag and drop them here and they will be classified automatically

13. | Build: Pre-annotated example The platform automatically pre-annotates documents based on pre-defined fields, which appear as underlined text. These underlines may not match at first but will improve as annotators review and correct them. Please note that pre-annotations cannot be deleted, only ignored.

14. | Build: Validated example

15. | Build: Data annotation & validation All fields that were pre-annotated incorrectly, or are missing, can be added/modified manually by highlighting the relevant span of text and selecting the correct field from the dropdown list.

16. | Build: Training a custom model Active learning works for custom document types as well. To pre-annotate custom document types: Upload one sample document and select the ‘ custom ’ document type Add all relevant fields from scratch Annotate the document manually Upload the rest of the documents (# of documents will be in the recommendations) Uploaded documents will be pre-labeled automatically, using the previously defined fields in step 2 DocPath LLM will make training custom models easier in the future

17. | Build: Best practices Always prefix your samples with the vendor/layout name and document type (if you are planning to work with multiple document types in the same project)  Upload at least 30 documents for each document type – ensure these are a representative sample of all documents Uploaded documents are automatically processed (uploaded, digitized, classified and annotated) – ensure you validate all classifications & pre-annotations and correct any mistakes  Mark the fields that are not present in a document as ‘missing’ Follow the recommended actions to build a good performing model Use the default UiPath Document OCR for Latin script languages, Hebrew or Arabic Use the Extended Languages OCR for Chinese, Japanese, Korean or other languages that the default OCR doesn’t support

18. Active Learning Best Practices Measure and Deploy

19. | Measure: Introduction The main objective of the Measure phase is to assess the model’s performance and identify areas for improvement. This includes the overall project’s performance, classification performance and extraction performance.

20. | Measure: Classification score Factors tab provides a list of recommended actions to improve model’s performance, including dataset size or document type(s) performance. Metrics tab provides key performance statistics for each document type, including: Train : number of documents on which the model was trained Test: number of documents on which the model was evaluated Precision: measures how often the positive predictions are correct Accuracy: measures how often the model correctly predicts the outcome Recall: measures the proportion of the total possible true positive results that the model was able to identify F1 score: harmonic mean of both precision and recall Y ou can double click on any recommended action in Factors, and it will bring you to the relevant training step in Build

21. | Measure: Classification score Classification score factors in the model’s performance as well as the dataset’s size and quality. It’s only available for projects with multiple document types. It’s comprised of 2 components: Factors and Metrics .

22. | Measure: Extraction score Factors tab provides a list of recommended actions to improve model’s performance, including number of uploaded/annotated documents, or fields accuracy for each document type. Dataset tab provides detailed information about documents used for training, total number of imported pages, and total number of labelled pages. Metrics tab provides key performance statistics for each document type, including: Training pages: number of pages on which the model was trained Rating: same as the project score rating (poor, average, good, excellent) Accuracy: measures how often the model correctly predicts the outcome Y ou can double click on any recommended action in Factors, and it will bring you to the relevant training step in Build More labelled training data is required More labelled training data is recommended Labelled training data target is achieved No fields created

23. | Measure: Extraction score Extraction score factors in the overall model’s performance as well as the dataset’s size and quality. Each document type has its own extraction score available. It is comprised of 3 components: Factors , Dataset and Metrics .

24. | Measure: Best practices You need to upload at least 10 documents to get a project score You need to upload at least 10 documents within the same document type to get a document type score You should stop training your model when it reaches your desired performance – this will vary per each use case and your business objectives Follow the recommended actions to fine-tune model’s performance

25. | Publish: Introduction The main objective of the Publish phase is to create and deploy new project versions to automate your processes. Create: freeze the current model’s state into a new project version Deploy : assign hardware resource and make the model accessible from workflows Automate : consume the product version in automations

26. | Publish: Creating model version Project version is a snapshot of the current model’s state. If you are happy with your model’s performance or want to test it in a workflow, create a new project version. Once a project version is created, model training is triggered automatically. The selected classifier and extractor models are trained on all uploaded data. This process can take a few minutes to several hours, depending on the model and size of the dataset. Click on the ’Create project version’ button Fill in the name (required) & description (optional) Choose the classifier and/or extractor models you want include Choose a deployment type Click ’Create’

27. | Publish: Deploying model version Once created, each project version will have a status. This can be one of the following: Training:  Models are being trained (temporary state) Trained:  Models are trained Undeployed:  Model training is complete, but the models are in idle state. Models are not consuming any resources and the project version cannot be referenced/used in workflows. Undeployed project versions are not available in Studio activities Deployed:  Model training is complete and is running. Models are consuming hardware resources, and the project version can be referenced from and used in workflows To deploy a project version, enable this toggle

28. | Publish: Best practices Each published version is referenced in your automations which consume the respective document types or classifier – if you wish to update the version used, make sure to update your automations You can easily do this by maintaining asset values in Orchestrator, or a mapping table (if you’re working with multiple project names and version) Make sure to select the required resources (classifier or document types) when creating a project version If using IntelligentOCR.Activities v6.22.0 and above, project names and versions can reference variables which can be maintained as Orchestrator assets, or within a mapping table if you are working with multiple projects in 1 process

29. | Which environment should I use to build, test and deploy Modern Document Understanding models? We recommend using a single project for development, testing and production. WHY? Governance Model versions RBAC Control which users have access to which projects* and who has permissions to publish or delete project versions DU Administrator has full access to perform any action DU Developer can read and manage projects but cannot create/delete projects DU Model Trainer can view projects, label documents, edit fields, import/export/delete data DU Viewer can view entities but cannot edit/delete them Centralize your automation resources within the project and manage all resource through project versions A project version contains a classifier and one or more extractors (or no extractors if the use case only requires classification) The new project architecture ensures that the Build phase, Active Learning training runs, creating and managing project versions, and consumption of these versions are completely independent and do not influence any aspect of run-time usage of deployed project versions Manage your project in a single place, including any sensitive data Test all your ML models and tools for OCR, classification, and data extraction in one place Move to production without incurring any retraining or moving configurations from one project to another Hide away the complexity of managing settings, ML models, deployment environments, replicas, etc *Can be set at both project-level as well as tenant-level

30. | Which environment should I use to build, test and deploy Modern Document Understanding models? We recommend using a single project for development, testing and production. WHY? Fine-tuning Other improvements Ability to tag versions as production or staging , and configure automations to use certain tags and document types, enabling a seamless transition from one version to another without the need to re-publish your automations Ability to configure activities to target projects from a different tenant or organization than the one the robot is connected to Ability to fine-tune the project using feedback from human validation in Action Center Data related to validated documents will be made available so that the project can be improved based on actual usage data *Please note that these features are still in development*

31. | Monitor: Project Performance (in preview) Project success metrics Consumption metrics* Runtime metrics Estimated time saved Estimated cost Processed documents by consumer Validation time Average handling time Straight-through processed documents Document types requiring validation Field corrections Extraction accuracy Classification accuracy Classification confusion Validation actions Exceptions AI units consumption overview AI units consumption details Runtime consumers  * Subject to change

32. | Monitor: Project Performance (in preview) The Project Performance tab displays an Insights dashboard with metrics to help evaluate the success of your project Project Performance metrics require  UiPath Insights to be enabled (no license required) It is currently in public preview , so feedback is highly encouraged

33. | Monitor: Processed Documents The Processed Documents tab provides a list of documents which have been processed in the selected project, either via APIs or Activities. Each document that has been digitized or processed will appear here. You can see the following information: File name Document type Consumer: API:  consumer which has digitized at least one page using the selected project ( Hover your mouse over this field to view the AppId to identify the consumer) RPA:  Studio Desktop or Studio Web projects containing at least one Document Understanding activity referencing the current project (Hover your mouse over this field to view the process name) Modified Date:  date when the last operation on the document occurred Validator:  username of a user who validated the task. If there is no validation task created for the respective document, N/A will be displayed AI Units:  number of AI Units consumed Click on the file name to get to the Document Details view. For more information on Document Details, see here .

34. | Monitor: Processed Documents The Processed Documents tab provides a list of documents which have been processed in the selected project, either via APIs or Activities. Each document that has been digitized or processed will appear here. You can see the following information:

35. Active Learning Best Practices Leverage on Hosting Benefits

36. Active Learning comes with a new training and deployment infrastructure that benefits from advanced algorithms for balancing training and model deployments. The new approach provides faster training times and more efficient model deployments and does not consume AI units for model training and model serving. The system automatically triggers training when necessary to provide a quicker feedback loop on the model's performance. Model deployments run on GPUs and scale automatically under load , so you don't need to configure the number of replicas or replica size. What this means : Allow for greater customization in segregating layouts that are more complex Allow for targeted training across layouts with different fields and naming conventions Ease of feedback loop for model health and redeployment | Model training and serving for Modern projects

37. Best approach to labelling is to upload and train the model layout by layout Keep a training record based on the scope of vendors/layout and update as training progresses Observe which layout decreases accuracy Keep a lookout for layouts that have complex labelling requirements (e.g. multi-line rows, column fields that cut across multiple columns in the table, etc.) | Segregating Layouts based on complexity and feedback from Active Learning

38. For more complex layouts, some column values have to be retrieved differently For example, Part Number could consist of different columns, best extracted in its constituents first before data transformation is applied to achieve a generic taxonomy Maintain the labelling and transformation logic in your design document | Achieve targeted training across layouts with different fields and naming conventions

39. Datasets can be easily exported and moved between projects – modern and classic alike. Export of data set can be done via Publish Always monitor the model health after training is complete for a certain layout Data sets can be manually modified in the zip files downloaded or after the data set has been uploaded into modern Steps to import a data set into Active Learning : | Ease of feedback loop for model health and redeployment Navigate to your Modern Project, click “Add new document type” button and create a new Custom document type. On the new custom document type is created, click “Upload” and select the zip file you exported from your other project. Wait for the upload to finish. 1 2 3

40. Active Learning Best Practices Overview of DocPath LLM - how it impacts the solutioning

41. Higher overall accuracy, particularly with tables. Reduced effort required for labeling data, thanks to improved ability to adapt to unknown document layouts. Less labeling effort is needed because there’s no longer a requirement to annotate every field instance on all pages where they may appear . Annotating one value per document is now sufficient. For example, if the Invoice Number appears on every page of a five-page invoice, not all five appearances need to be annotated. With this Private Preview, annotating it once, on any page, is all that's required. Increased automation rates (Straight-through processing rate) are due to the enhanced correlation between confidence level and accuracy . With a given confidence threshold, fewer documents will be directed to the Action Center for the same level of accuracy. It is essential to understand that the names of the fields now impact the model's performance . Field names should be written in natural language with proper grammar and should only use widely known acronyms such as Number/No, Account/Acct, Address/Addr, and Apartment/Apt. Currently, only West European languages are supported, so field names should also be in West European languages. Please avoid non-descriptive names like Column 3 unless that is how they are referred to in the document itself . DocPath LLM and its intended benefits

42. Active Learning Best Practices Consuming your model in StudioX or Studio

43. Generative Document Understanding Activities

44. Minimal setup Shines when dealing with unstructured data Flexibility Why Gen AI For Document Processing?

45. Document Understanding: Generative Classification What is it? Document classification made easy with Generative AI Classifying documents is fast and easy with Gen AI – just define the document types, no need to write rules or train new ML models

46. Document Understanding: Generative Extraction What is it? Question-answering model powered by Generative AI Generative AI can answer questions and summarize content which works perfectly for free-form unstructured documents – with no need to train custom ML models.

47. Document Understanding: Generative Validation What is it? Get a ‘second opinion’ on the extracted data from Generative AI to reduce the human validation effort With Generative AI used to confirm the extraction output, the overall automation rate can increases by up to 200% and the average handle time decreases – reducing the time spent on human validation. Source: Test by UiPath AI R&D on a diverse set of enterprise documents​

48. Prompt Engineering For classifier For extractor Use precise language provide as much context as possible, a detailed description of each document type or each fields

49. GenAI’s strength lies in unstructured document types An example Letter Result

50. Demo

51. Q&A

52. Thank you!