3. 2023
Our goal is to disrupt the experience
and financial services we provide to our
customers through GenAI.
4. Millions
Billion
few millions conversational responses
per month.
few billion tokens sent to Azure OpenAI
per month.
+45%
NPS.
+50%
resolution-oriented interactions
in the first contact.
Assistant in numbers
5. Assistant, customer service case
Conversational Generative AI
solution-oriented, simplifying and personalized conversation for all customer
6. 1
2
3
4
5
Conversational
Being able to interact in natural language and guide users
through a process of discovery while learning and
remembering new relevant information along the way.
Domain Knowledge
Understanding the entities, terminology, categories, and
attributes of each specific use case, as well as leveraging
quality statistics on typed terms.
Predictive
Being able to provide the answer to the question in
advance, understanding what users mean.
Simplifier
Bringing the solution to the task, with direct delivery
of the sought-after information or action.
Personalized
Being able to use user context, question context,
and domain context to better understand the user's
intention.
5 pillars that support our solution
9. Virtual
Assistant
Cutting-edge Generative
AI technology
Channel on Direct
Message
Integration with several
PicPay APIs
Simplified and supported
transfer
API integration
platform
Article
management
Contextualized
deeplinks
Performance metrics and
insights using GenAI
10. AI Assistant
Taxonomy products
and categories
User sends a
message
The assistant
provides an answer
API Discovery
Engine
Conversational
Engine
Knowledge Base
User Context
Engine
Integrations with
internal services
APIs
Intent classifier
Answer
Generator
FAQs, resolution conversations,
and deeplinks list are processed
and indexed in the engine.
AI model trained to classify
intentions
Human Assistant
prompt
Inserting
instructions in
the prompt
Escalation
pass control
Architecture
Deeplink Builder
(Call to action)
Security
Security
11. RAG, Retrieval-Augmented Generation
No model retraining
Cost-effectiveness
Data up to dated
Human review of created articles
Experience with a Information Retrieval
12. RAG
RAG - Retrieval Augmented Generation
Semantic
Search
Similarity Search
(Vector/Embeddings) +
BM25
Conversation
Engine
Control over content and Prompt Engineering
Focus on PicPay and not on the internet content
Provide up to date knowledge without needing to retreaning LLM Models.
LLM Model
Prompt
Answer
Generator
FAQs, resolution conversations,
and deeplinks list.
Human Moderation.
API Discovery
Engine
Cartão Virtual
14. GenAI with Azure OpenAI
Data security
Multi-cloud solution
Reduced Latency
Brazilian support
Partnership
15. Tips. Lessons takeaways
Be specific in the prompt
RAG or Fine Tuning depends on the use case
Control the flow of token consumption
People who act end to end
Be prepared that everything will change