Vertesia Prep Flashcards
(26 cards)
Describe Embeddings
Technical Definition:
- Embeddings are high-dimensional vectors (sets of numbers) representing tokens (words).
- They capture the semantic meaning and relationships between words.
Non-Technical Definition:
- Embeddings are like putting words onto a map, where each word’s location shows its meaning and how it relates to others.
- The computer learns where to put each word on the map by studying lots of language.
- Words with similar meanings end up close together on the map.
- This helps AI understand language and make sense of how words connect.
Explain why tokens are used for models and not just entire words.
Words are made up of smaller building blocks. By working with those smaller pieces, the AI can understand new or unusual words more easily—like “unhappiness,” which it can split into “un,” “happy,” and “ness.”
If the AI only looked at whole words, it would be confused by any new or misspelled word it hadn’t seen before.
Breaking sentences into smaller pieces is like cutting a puzzle into more pieces: you can fit new shapes together, even if you’ve never seen them before.
This way, the AI can “read” and “write” almost anything, even weird words or made-up names!
What are the 2 distinct sets of API available through the Vertesia platform?
Studio API
manage projects, environments, interactions, prompts, and more
JWT to authenticate (api key can be used for API requests if needed).
Zeno API
manage content objects, content types, workflows, and more
JWT to authenticate (no API key authent available).
What 5 use cases was Vertesia CLI designed for?
- List and switch between your Vertesia projects
- List the existing interactions and execution environments
- Run interactions once or multiple times over a set of different data inputs
- Generate data inputs to run the interactions against
- Search through the history of runs to inspect detailed results
Difference between Authorization and Authentication
Authentication is proving who you are.
* username and password
* fingerprint or face id
* security tokens (sms codes and authenticator apps).
Authorization is controling what you can do
* RBAC (role based access control)
* ACLs (access control lists)
* Attribute based access (based on user traits like dept)
Authentication is showing ID at door, and Authorization is getting past velvet rope into VIP
Discuss Firebase and OIDC
Firebase is google’s backend-as-a-service platform from google.
* Integrate firebase auth into an app, and then when a user logs in firebase authenticates them, then passes app a user object
* authorization with firebase is handled through Firebase Security Rules
OIDC is an industry standard protocol that sits on top of OAuth 2.0
* is about providing proof of identiy, usually with JWT tokens
Explain NoSQL, Graph Search, Semantic Search, and Structured Search
NoSQL:
A flexible, scalable database type for storing unstructured or semi-structured data (like JSON documents). Used mainly for fast storage and basic search/filtering.
Graph Search:
A method of searching data that’s organized as nodes and relationships (graphs), ideal for finding connections and paths between entities (e.g., social networks, knowledge graphs).
Semantic Search:
Search based on meaning and context, not just exact words. Uses embeddings or LLMs to find results that are similar in intent, even if the keywords don’t match.
Structured Search:
Traditional search over well-defined, tabular data (like SQL databases). Enables precise queries and filtering on specific fields and values.
What does Stateless mean in software?
Each request or interaction is independent and contains all the information needed to understand and process it—no memory of previous requests is kept by the server or service.
Importance: Scalable, simpler, more reliable (if one server fails another can handle the request).
Opposite would be “stateful”: server remembers past inteeractions and sessions - it keeps “states” about users.
What is the difference between APIs and Web Hooks?
Webhook is a type of integration that uses APIs, but not all APIs are web hooks.
Webhooks are event driven and push data
APIs are request-driven and pull data
Examples:
* Webhook: company is notified when a new user signs up
* API: user logs in and request pulls all of their info for them to see in the app
Explain Message Queues and Event Streams along with popular tools for this.
Message Queues:
* One app messages another and it’s stored in a queue until receiving app can process (single use pickup locker for packages)
* Tools: RabbitMQ, AWS SQS, Azure Queue Storage
* Use cases: decoupling systems (order processing or email delivery)
Event Streams
* Designed for continues high-volum, real-time flows of data that can be processed by multiple consumers at once (like a bulletin board with multi-people reading it).
* Tools: Apache Kafka, AWS Kinesis, Azure Event Hubs
* Use Cases: real time analytics (IOT), Feeding data pipelines (log processing, fraud detection).
What is iPaas
A cloud-based tool that makes it easy to connect different apps and data sources—often with simple drag-and-drop workflows—so information can move automatically between systems without custom coding.
Zapier, Mulesoft, Boomi, Workato, Tray.io
Compare and Contrast RAG vs. Fine Tuning
Both are methods used to improve the accuracy of LLMs
Fine Tuning
* take a trained model and train it more on your own data.
* Like an employee with experience starting a new job. They have base knowledge, but need training on specifics.
* Best for stuff that doesn’t change and when you don’t need the model to look stuff up before answering.
RAG
* Model looks up relavent info before responding and uses that info to guide it’s response.
* Employee checking their documentation before answering a clients question
* Best if you’re referencing content that is subject to change.
Explain Zero-shot, One-shot, and Few-shot prompting.
Refers to the number of examples given to a model.
Zero-shot is best when task is clear
One-shot is best when a single example will suffice
Few-shot is best when a model needs to understand a pattern or rule with examples
Explain Prompt-chaining
A technique where you break a complex task into a series of smaller prompts, passing the output of one prompt as the input to the next. This helps the model handle multi-step reasoning or workflows.
Example: Put a summary of something in an email. First summarize, then draft email.
Vertesia has the architecture to support __________________
- Multiple teams
- Use cases
- Models
- Inference providers
- Compliance requirements
List 11 important features that are handled by Vertesia meaning companies would have to build these on their own if they went a custom route.
- Governance
- Security
- Orchestration
- Virtualization
- Fine-tuning
- Workflows
- Agentic RAG
- integratoin and deployment options
- Collaboration
- analytics and monitoring
Describe how governance and security are handled in Vertesia
Enterprise-grade certifications: SOC 2 Type II & ISO 27001.
Encryption: Data is encrypted at rest and in transit.
Role-based access & SSO: Granular permissions plus SSO (SAML, OAuth2, OpenID).
Comprehensive audit logging: All user actions and workflows are logged for traceability.
Policy-based guardrails: Enforce and validate safe AI inputs/outputs.
Data privacy: Your data is never used to train third-party models.
Customizable data retention: Fine-tuned retention settings for compliance.
Describe how virtualization is handled in Vertesia and the benefits it offers
Virtualization in Vertesia refers to the platform’s ability to create a “virtual” large language model (LLM) by seamlessly connecting multiple underlying AI models and providers together.
Instead of being tied to a single AI provider (like OpenAI or Google), Vertesia lets you treat a group of models as a single, unified resource.
When you send a request to this virtual model, Vertesia automatically decides which underlying provider or model should handle the task—based on your preferences for cost, speed, reliability, or quality.
Describe how orchestration is handled in the Vertesia platform
V has a centralized workflow orchestration engine built on Temporal. It ensures that your tasks run reliably across multiple models and systems. This scales horizontally.
You can chain together multiple GenAI tasks that follow business logic and can interact with internal or external tools.
Every workflow is logged and visible in the platform which allows for better governance.
Describe how fine-tuning is handled in Vertesia and some use cases for it
tbd
Describe how workflows work in Vertesia
tbd
Describe Agentic RAG as it relates to Vertesia
tbd
What kind of analytics and monitoring does Vertesia offer
tbd
How does Vertesia handle collaboration among teams in their platform?
tbd
