H2O Generative AI Starter Track - Support Presentation Slides.pdf

H2O.ai Conﬁdential
H2O Generative AI Starter
Track
Fundamentals and Applications
with h2oGPTe

● Understand the foundations of Generative AI and
related concepts
● How to use Generative AI on your data (documents,
audio, images etc;)
● How to conﬁgure RAG and use prompt templates
Objectives & Learning Outcomes

● Three AI Paradigms: Machine Learning, Deep Learning, and Generative AI.
● Large Language Models (LLMs): Power Generative AI with accurate,
context-aware outputs.
● Applications: AI-generated visuals, audio, and automated communication.
● Prompt Engineering: Techniques to guide AI for precise results.
● Retrieval-Augmented Generation (RAG): Combines retrieval and generation
for context-rich responses.
● Next Steps: Apply these tools with H2O Enterprise GPTe for tailored AI
solutions.
Recap: Fundamentals of Generative AI

Part 1
Fundamentals of GenAI
● AI / ML
● Gen AI
● LLM
● Prompt
● RAG

Foundations of GenAI
● AI / ML
● Gen AI
● LLM
● Prompt
● RAG
Machine
Learning:
Deep
Learning:
GenAI:
Model
(e.g. Gradient Boosted
Machines / Linear
Models)
Model
(e.g. CNNs,
Transformers)
Model
(e.g. Large
Transformers)
Numbers Numbers
Multimedia Numbers
Multimedia Multimedia
“Predict hypertension
risk using age, gender,
physical activity, etc.”
“Detect if a patient has
COVID-19 from chest
X-ray image”
“Perform patient
screening with
interactive chatbot”

● AI / ML
● Gen AI
● LLM
● Prompt
● RAG
AI that generates new content based on its learnings
from the existing content
Process of learning from data is called “training” that produces a “model”
Very large model trained on very large text data on
large compute is an LLM
Produces a highly cohesive text with correct structure, and meanings
GenAI
LLM

● AI / ML
● Gen AI
● LLM
● Prompt
● RAG

● AI / ML
● Gen AI
● LLM
● Prompt
● RAG
Prompt is the input
to a Gen AI System
or to a LLM
A nicely crafted
prompt can yield
better and relevant
outputs, A bad
prompt may not
User’s prompt
LLM Response

● AI / ML
● Gen AI
● LLM
● Prompt
● RAG
Role Play
○ “Act as a social media inﬂuencer…”
○ “Act as an expert in astrophysics…”
Zero Shot Prompting (not providing any examples)
○ Classify my text into positive, negative, neutral
Provide Instructions (guardrails) with Prompt
○ “Please use my document for the answer, here is the link : <>
○ “Share all the important ﬁnancial metrics from the report, but do not
use any sensitive information”
Prompt Chaining
○ “You are a healthcare recruiter. You’re good at writing interview
questions. Please ask me each question below one at a time”

● AI / ML
● Gen AI
● LLM
● Prompt
● RAG
Retrieval Augmented
Generation (RAG)
Retriever + Generator
● Retriever: retrieves relevant information
from the documents (context)
● Generator: LLM that generates accurate
response based on the retrieved context
● Workﬂow
Ingest → Parsing → Chunking → Indexing
→ Embeddings* → VectorDB*
● Inference
Prompt → Relevant Chunk Extraction →
LLM → Response

Part 2
Use GenAI on your Data
h2oGPTe

1. Prompting (Chat)
Use GenAI on your Documents: h2oGPTe
Enterprise Generative AI Platform
to perform Question Answering on
your documents (or websites,
webpages, and workplace content -
in an internet disconnected setup)
2. Data Ingestion (Collections)
3. Conﬁgure (Customize)
4. RAG (Document QA)
5. API

Potential Use Cases
Task Benefit
Generate a first cut draft of a report on the
terrorism situation in Southeast Asia using
uploaded news articles
The traditional process is tedious and takes time.
Enterprise GPTe can help officers save time with a
good enough first-cut draft for further refinement.
Draft a first cut speech on organisational
development
Help answer queries on HR or procurement policy Officers can save time on fielding replies since
their colleagues can now self-service with the
chatbot.

Free Platform Access for
training purposes
https://h2ogpte.genai.h2o.ai/
Step 1:
Go to
Step 2:
Sign in with your
Google or Github
account.
You will be directed to the
id.public.h2o.ai hyperlink

Step 3:
Once in, you are free to
explore our public
Enterprise h2oGPTe
Enjoy!
Free Platform Access for
training purposes

1. Prompting 2. Collection 3. Conﬁgure 4. RAG 5. API
Chat with the LLMs
● Who are you?
● How are you doing?
● What are the factors of drug desistance?
Recommended PDF File to work with

Ingest Anything:
documents, audio,
images, and websites,
over 65 different ﬁle
formats → Group of
related documents is a
COLLECTION
* Important! Only upload
training documents

Chat Configurations
LLM
● Automatic mode: selects the optimal
LLM for the job from all the configured
models
● Or select a specific language model
Generation approach
● LLM Only: does not use any documents
to generate an output
● RAG (Retrieval Augmented Generation):
uses retrieval-based systems and
generative models to produce more
accurate and contextually relevant
responses

Prompt Template Configuration
A prompt template is used to guide the model's
behavior and influence the style, tone, and
content of its responses.
Some key functions of a prompt template
include:
● Helping the model understand the task it
needs to address
● Setting the desired tone and style of the
response
● Simulating specific roles (e.g. customer
service agent, technical expert)

Ask Questions on the
documents
- “What are the factors of drug
desistance”
Ask Questions on Image
- “What helps desisters stay away
from drugs?”
Ask Questions on Audio
- “summarize the ﬁle”

To generate a new API key

BONUS: Creating and sharing your first prompt template!
But first, what is the difference between System Prompt and User Input
● Predefined instructions or context
● Primarily to guide LLM on how to behave or
respond (think: role / persona)
● Sets the following:
○ Tone
○ Rules
○ Structure
○ Task
○ …
● Generally not visible to user
● Message or question entered by user during
an interaction with the LLM
● Could be a question, request or feedback
etc.
● Dynamic and changes with each interaction
(unlike system prompt that stays unchanged)
System Prompt User Input

System Prompt Template for Report Generation
#IDENTITY and PURPOSE
You are an expert at generating high-quality, detailed reports based on the input provided.
Your goal is to take the user’s input topic, analyze it thoroughly and generate a well-structured,
comprehensive report with actionable insights.
Follow the steps below to generate a report that is clear and provides deep insights into the topic.
#STEPS
- First, carefully read and understand the user’s input topic and identify the key themes and
objectives for the report.
- Create an outline for the report. Breakdown the report into sections according to the user’s
input topic such as,
Introduction
Analysis
Recommendations
Conclusion
- Imagine yourself as an expert in the given input topic and write the report from the
perspective of someone with deep knowledge in that ﬁeld.
- Provide facts, relevant data and examples to support your analysis and also use concise and
clear language to explain complex topics.
- Maintain a professional and neutral tone throughout the report.
- If you have data or examples based on the input topic, make sure to use them effectively to
demonstrate the concepts or arguments.
- Structure the report with clear headings and sub-headings and use bullet points, numbered
lists and visuals when necessary.
- The report should be between 300-500 words, depending on the complexity of the topic.
#Output Guidance
- The report should contain a clear introduction, well-structured body sections and a concise
conclusion.
- Ensure that the report addresses the user’s input topic in depth.
#Output instructions
- Output the report by providing a section for each part of the report - Introduction, Analysis,
Recommendations, Conclusion.
- The INTRODUCTION section should briefly explain the topic and its relevance and state the main
objective of the report.
- The ANALYSIS section should go deeper into the subject by breaking down key components and
use relevant data, facts and examples to back up the analysis.
- The RECOMMENDATIONS section should provide actionable and specific recommendations
based on the analysis.
- The CONCLUSION section should summarize the main points discussed in the report.

#IDENTITY and PURPOSE
You are an expert at generating high-quality, detailed reports based on the input provided.
Your goal is to take the user’s input topic, analyze it thoroughly and generate a well-structured,
comprehensive report with actionable insights.
Follow the steps below to generate a report that is clear and provides deep insights into the topic.
#STEPS
- First, carefully read and understand the user’s input topic and identify the key themes and
objectives for the report.
- Create an outline for the report. Breakdown the report into sections according to the user’s
input topic such as,
Introduction
Analysis
Recommendations
Conclusion
- Imagine yourself as an expert in the given input topic and write the report from the
perspective of someone with deep knowledge in that ﬁeld.
- Provide facts, relevant data and examples to support your analysis and also use concise and
clear language to explain complex topics.
- Maintain a professional and neutral tone throughout the report.
- If you have data or examples based on the input topic, make sure to use them effectively to
demonstrate the concepts or arguments.
- Structure the report with clear headings and sub-headings and use bullet points, numbered
lists and visuals when necessary.
- The report should be between 300-500 words, depending on the complexity of the topic.
Deﬁning Role and Purpose
Chain-of-thought reasoning
Detailed structure guidance
Tone and Style
Role-Playing as an Expert

#Output Guidance
- The report should contain a clear introduction, well-structured body sections and a concise
conclusion.
- Ensure that the report addresses the user’s input topic in depth.
#Output instructions
- Output the report by providing a section for each part of the report - Introduction, Analysis,
Recommendations, Conclusion.
- The INTRODUCTION section should briefly explain the topic and its relevance and state the main
objective of the report.
- The ANALYSIS section should go deeper into the subject by breaking down key components and
use relevant data, facts and examples to back up the analysis.
- The RECOMMENDATIONS section should provide actionable and specific recommendations
based on the analysis.
- The CONCLUSION section should summarize the main points discussed in the report.
Output instructions for coherent response

Now, try it out!
Draft a 300-word report on the criminality issues associated with drug desistance.

Agent Approach
29
Multi-Model Integration Code First Agent
Minimalist Approach
Combines standard models for
audio-visual understanding and
reasoning, paired with Sonnet 3.5 for
natural language processing and
reasoning tasks.
CodeAct (Code + Act) framework
extends code generation and
execution as the primary form of
action
Single agent with minimal scaffolding,
avoids specialized methods or
complex orchestration pipelines
Uses raw model without additional
ﬁne tuning

Example Workflow
30
Analyze the stock
price of Tesla
Prompt
Load stock price from
yfinance in a python
notebook
Code
Execution
Error loading yfinance,
switching to
alphavantage api
Error
Handling
Execute
google_search.py with
the modified prompt
Agent tool:
Websearch
Utilize BeautifulSoup
to scrape webpages
for additional details
Code
Execution
Agent Analysis

Summary
Use GenAI on your Documents with h2oGPTe
- Prompting (Chat)
- Use prompting techniques to interact with LLMs to get desired response
- Prompt templates can be created and shared
- Data Ingestion (Collections)
- Ingest various data types (documents, audio, images, websites)
- Supports over 65 ﬁle formats grouped into “Collections” for easy management
- Retrieval-Augmented Generation
- A technique used to ensure contextualised response by providing relevant documents for LLM to retrieve data
from - improves model’s response accuracy
- Conﬁguration
- Customise settings such as the selected LLM, RAG approach, and prompt template
- API
- Allow technical users to build applications around h2oGPTe via API

1.
Prompting
2.
Collection
3.
Conﬁgure
4. RAG 5. API
Explore h2oGPTe in Action – 5 Steps

SCAN THE FOLLOWING QR CODE TO ACCESS THE PDF
Drug_use_desistance.pdf

WHAT ARE APIs?
- APIs act as a bridge between software applications, enabling seamless communication.
- They allow you to send
requests (e.g., ask
questions, retrieve
insights, generate
content).
- Responses are
provided in a
structured,
programmatic way for
easy integration.
- APIs enable
embedding h2oGPTe's
capabilities directly
into your systems,
beyond the standard
interface.
H2O Generative AI Starter Track
Slide 28.1 b4

TYPES OF API KEYS
H2O Generative AI Starter Track
Slide 28.1 b4
🔑 Global API Keys –
Grant full access, allowing API calls to impersonate the
user.
📁 Collection-Speciﬁc Keys –
Restrict access to a particular collection for controlled
usage.

H2O Generative AI Starter Track - Support Presentation Slides.pdf

More Related Content

Similar to H2O Generative AI Starter Track - Support Presentation Slides.pdf (20)

More from Sri Ambati (17)

Recently uploaded (20)

H2O Generative AI Starter Track - Support Presentation Slides.pdf