Dev Dives: Train smarter, not harder – active learning and UiPath LLMs for document processing

💥 Speed, accuracy, and scaling – discover the superpowers of GenAI in action with UiPath Document Understanding and Communications Mining™:

See how to accelerate model training and optimize model performance with active learning
Learn about the latest enhancements to out-of-the-box document processing – with little to no training required
Get an exclusive demo of the new family of UiPath LLMs – GenAI models specialized for processing different types of documents and messages
This is a hands-on session specifically designed for automation developers and AI enthusiasts seeking to enhance their knowledge in leveraging the latest intelligent document processing capabilities offered by UiPath.

Speakers:
👨‍🏫 Andras Palfi, Senior Product Manager, UiPath
👩‍🏫 Lenka Dulovicova, Product Program Manager, UiPath

1. 1
Train smarter, not harder –
active learning and UiPath
LLMs for document
processing
UiPath Dev Dives Webinar Series

2. 2
Safe Harbor​
This presentation may include forward-looking statements. Forward looking statements include all statements that are not historical facts, and in some cases, can be
identified by termssuch as “anticipate,” “believe,” “estimate,” “expect,” “intend,” “may,” “might,” “plan,” “project,” “will,” “would,” “should,” “could,” “can,” “predict,”
“potential,” “continue,” or the negativeof these terms, and similar expressions that concern our expectations, future performance, strategy, estimates of market size and
opportunity, plans or intentions. By their nature, thesestatements are subject to numerous risks and uncertainties, including factors beyond our control, that could cause
actual results, performance or achievement to differ materially andadversely from those anticipated or implied in the statements. These and other risk factors are
described in the “Risk Factors” section of our Annual Report on Form 10-K filed annually with theSecurities and Exchange Commission following the conclusion of our
fiscal year ended January 31as well as in our Forms 10-Q and otherfilings withthe Securities and Exchange Commission. Althoughour management believes that the
expectations reflected in our statements are reasonable, we cannot guarantee that the future results, levels of activity, performance or events andcircumstances
described in the forward-looking statements will be achieved or occur. Recipients are cautioned not to place undue reliance on these forward-looking statements,
whichspeak only as of the date such statements are made and should not be construed as statements of fact. Except as may be required under the federal securities
laws, we undertake noobligation to update these forward-looking statements to reflect events or circumstances after the date hereof, or to reflect the occurrence of
unanticipated events.​
[Certain information contained in this presentation and statements made orally during this presentation relate to or are based on studies, publications, surveys and other
data obtainedfrom third-party sources and UiPath’s own internal estimates and research. While UiPath believes these third-party studies, publications, surveys and other
data to be reliable as of thedate of this presentation, it has not independently verified, and makes no representations as to the adequacy, fairness, accuracy or
completeness of, any information obtained fromthird-party sources. In addition, no independent source has evaluated the reasonableness or accuracy of UiPath’s internal
estimates orresearchand no reliance should be made on anyinformation or statements made in this presentation relating to or based on such internal estimates and
research]

3. 3
Meet today’s speakers:
LenkaDulovicova
Product program manager
UiPath
AndrasPalfi
Senior product manager
UiPath

4. 4
Agenda
01
02
03
Introduction to IDP
UiPath CommPath & DocPathLLMs
04
Product demo
Active Learning in Document Understanding
05
Q&A

5. 5
The enterprise is inundated with documentsprocessed manually
leading to business inefficiency
Finance
•Invoices
•Purchase orders
•Expense reports
HR & People
•Candidate applications
•Onboarding documents
Document processing challenges
Limit business growth and scalability
Labor-intensive document processing limits the ability to
scale efficiently and capture market opportunity.
Result in poor customer experience
Complex, unstructured data mandates human
decisioning, slow onboarding, and servicing.
Increase risk
Higher chance of data input errors, missed information,
and incorrect procedures.
Sales
•Contract agreements
•Orderamendments
Customerqueries
•Customer emails
•Customer tickets

6. 6
Business runs on documents and communications
Enterprise
Healthcare
Banking & Financial Services
Public Sector
Insurance
Manufacturing
Hiring & onboarding processes
Customer service & support processes
Finance processes (e.g. accounts payable & receivable)
Sales & order management processes
KYC – Know Your Customer
Trade settlement & amendment processes
Mortgage application & processing
Client & customer onboarding
KYC – Know Your Customer
Underwriting processes
Claims handling & processing
Patient registration & processing
Consent and assent processes
Health insurance claims
Immigration application & processing
Benefit application &processing
Unemployment verification & processing
Sales & order management processes
Accounts payable & accounts receivable
Purchase order processing
Customer & vendor communications
Customer queries &communications
Citizen questions & feedback Health history

7. 7
1 2 3
Understand ActReceive
End-to-end intelligent document processing (IDP) solution
Extracts relevant data from the
documents
Requests or
communications with
attacheddocuments:
•Multiple languages
•Various formats
•Handwriting
•Signatures
•Skewed & low-quality scans
•Checkboxes
•Tables
Extracts key intent, sentiment and
context data from messages
Human in the loop
Asking employees to validate the
results if required or in case of
inaccuracies and exceptions.
UiPath Automation
Route the extracted actions and data
to downstream systems for further
processing.

8. 8
UiPath uniquely brings AI + automation together for a seamless
end-to-end platform
Recognized
IDP leader
UiPath is a leader in product,
vision, strategy, and market
impact across multiple
categories in respected analyst
firms including Everest, Gartner,
Forrester, and more.
Embedded
Generative AI
Generative AI, paired with
Specialized AI to reduce
training and validation time,
providing faster time to value.
UiPath Business
Automation Platform
IDP as part of theleading
enterprise automation platform
that connects hundreds of data
sources and enables taking
actions at scale.
Documents
& Communications
70+ pre-builtmodels to analyze
and process different types of
documents and communications
across industries and domains.

9. Active Learning
The UiPath word mark, logos, and robots are registered trademarks owned by UiPath, Inc. and its affiliates. ©2023 UiPath. All rights reserved.

10. 10
General availability now
Active
Learning
Active learning employs an iterative process
between annotators and model to reduce the
amount of data required to train an ML model.
•80% faster model training
•Reduction of data samples for annotation
•Guided training experience (no ML or coding
skills needed)
•Built-in model performance data
General availability now
Active
Learning
ML model
Human
annotation
Evaluate
performance
Retrain and
redeploy
How does it work?
Training smarter, not harder.

11. 11
Active Learning in Modern Projects
Build
Load samples, annotate, and
train your model with a guided
experience
Measure
Understand your model
performance
and how to improve it
Publish
Deploy and manage
your projects
with ease
Monitor
Monitor and audit
the performance of
your automation
Active Learning brings a new unified experience to training, deploying, and monitoring Document Understanding models
in one user interface. This was previously across Document Understanding, Document Manager and AI Center.

12. 12
Modern Projects
•Automation Cloud only
•Generative pre-annotation
•Faster training time & more efficient model
deployment
•1 page = 1 AI unit & no infrastructure costs
•No document splitting available (yet)
•No auto-fine-tuning support (yet)
•Not compatible with IntelligentOCR Activities (yet)
•Consuming projects / moving datasets between
different tenants or organisations not possible (yet)
Key considerations:

13. UiPath LLMs
The UiPath word mark, logos, and robots are registered trademarks owned by UiPath, Inc. and its affiliates. ©2023 UiPath. All rights reserved.

14. 14
UiPath LLMs
Increased accuracy | Advanced unstructured data processing | Accelerated time to value | Robust security
LLM trained to process any document
OOTB with little to no training required,
including free-form unstructured data
and tables. This is an underlying LLM
that will be used to drive various
Document Understanding features.
UiPath DocPath
LLM trained to process communications
of varying complexity, including multiple
requests and fields and the relationships
between them. This is an underlying LLM
used to drive specific features in
Communications Mining, starting with
generative extraction.
UiPath CommPath

15. 15
Generative Extraction
powered by CommPath
Generative extraction in Communications
Mining is the use of AI and natural
language queries to understand and
accurately extract data from a message.
•Automate more:increase automation
rate,extract complex data, and understand the
requests, fields and the relationsbetween them
•Lower training effort: 2-3xless training effort for
the same levels of accuracy compared to
conventional training
•Fine-tune: CommPath can be fine-tuned to
extract specific fields and drive performance
General availability soon

16. 16
Out-of-the-box extraction
powered by DocPath
A generative large language model
specialized for a large and diverse set of
enterprise documents.
•Little to no training time: up to 10x training effort
reduction, with many use cases requiring no
annotation
•More accurate extraction: 45-76% error rate
reduction compared to market leading Generative
AI models
•Complex tables: 30-65% less errors compared
toother IDP and Generative AI vendors
General availability now
Phase 1: Use public endpoints for out-of-the-box extraction (not
retrainable):
General availability now for most of the document types
supported out-of-the-box:
-Purchase Orders, Remittance Advices,Bank Statements, Utility Bills, and 30+
more document types
Timeline for outstanding items:
-Invoices, Receipts, and IRS tax forms (1040x, 941x, 709, etc.) – coming soon
-Invoices Japan, China and Hebrew – on the roadmap
Phase 2: Build modern projects for your own document types
Timeline: Public Preview in 2024.10. We will also run a Private Preview soon,
which will be announced on the Insider Portal.

17. 17
UiPath DocPath&
UiPath CommPath
Built for enterprise automation. Robust
security, compliance and governance
paired withcomprehensivecontrols
and guardrails.
•Enterprise control:superior controls and
safeguardscompared to standard generative
LLMprocessing
•Standardize output:specify and normalize
necessary output into structured, usable
formats for automation
•Optimizeperformance:can be used in
unattended automations due to confidence
estimates, versioning, precision vs. recall
balance
•Data privacy: hosted by UiPath, so no
additional data sharing implications

18. 18
Which new scenarios will be enabled by DocPath and CommPath?
•DocPath:
•Once DocPath is available for custom model training with active learning, any structured and semi-structured documents should be covered out-of-the-box with little to no
training required
•Free-form unstructured documents like contracts
•Documents with complex tables (nested tables, tables where a line item extends onto a second page, merged cells, multi-column documents like transcripts/menus)
•Documents with long line items or groups of charges like utility bills or phone bills
•CommPath:
•Complex messages with multiple intents that need to be extracted as separate requests.
Which languages are supported by DocPath and CommPath?
•DocPath: Same languages as supported now by out-of-the-box Document Understanding models. Japanese, Chinese, right-to-left languages like Hebrew coming later.
•CommPath: Currently GA languages in Communications Mining including Japanese, Korean
How do we ensure enterprise controls and guardrails with UiPath LLMs?
•Superior enterprise controls and safeguardscompared to standard generative LLMprocessing(including performance evaluation and validation, confidence thresholds, customer-
specific fine-tuning, RBAC similar to other capabilities for all model operation activities).
•Standardized output- you can directly interact with the model to grant a structured output in a specific format meaning lower error rates.
•The LLMs are hosted by UiPath, so no additional data sharing implications.
•The models provide confidence estimates, versioning, etc., so can be used in unattended automations.
Common questions on UiPath LLMs

19. Demo
The UiPath word mark, logos, and robots are registered trademarks owned by UiPath, Inc. and its affiliates. ©2023 UiPath. All rights reserved.

20. 20
Useful resources
UiPath Documentation:
•Communication Mining docs -docs.uipath.com/communications-mining
•Document Understanding docs -docs.uipath.com/document-understanding
Exclusive assets for those who joined live:
•Active learning user guide
•FAQ on UiPath LLMs
Free trial available at uipath.com
Connect with Andras and Lenka on LinkedIn:
https://www.linkedin.com/in/andpal/
https://www.linkedin.com/in/lenka-dulovicova-1b081188/
Join the next Dev Dives sessions:
https://bit.ly/Dev_Dives_2024

21. 21
Live Q&A