AI Tools & Technologies Flashcards

What is 'context window' in the context of large language models?

Answer: The maximum number of tokens a model can process in a single input-output exchange

The context window defines the total token limit (input + output) the model can handle in one interaction, constraining how much text it can consider at once.

A client needs to generate SQL queries from natural language using an LLM. Which technique best improves accuracy by providing example query pairs in the prompt?

Answer: Few-shot prompting

Few-shot prompting includes example input-output pairs (natural language → SQL) in the prompt, guiding the model to match the expected format and logic.

Which tool allows AI consultants to version control and track experiments, hyperparameters, and metrics across ML model training runs?

Answer: MLflow

MLflow is an open-source platform for tracking ML experiments, versioning models, and managing the full ML lifecycle.

In agentic AI systems, what is 'ReAct' (Reasoning + Acting)?

Answer: A prompting framework where the model interleaves reasoning traces and tool-use actions iteratively

ReAct is a prompting paradigm that has the model alternate between reasoning (Thought) and taking actions (Act), enabling dynamic, iterative problem-solving with tools.

Which embedding model family from OpenAI is commonly used to convert text into dense vectors for semantic search and RAG?

Answer: text-embedding-ada / text-embedding-3

OpenAI's text-embedding models (like text-embedding-3-small and text-embedding-3-large) convert text into dense vector representations used for semantic search and RAG.

What problem does 'guardrails' software (e.g., Guardrails AI, NeMo Guardrails) solve in production LLM applications?

Answer: It enforces output constraints, validates structure, and filters harmful or off-topic content

Guardrails frameworks intercept LLM inputs and outputs to enforce content policies, validate schemas, and prevent harmful, biased, or off-topic responses in production.

Which approach enables an LLM to use real-time internet data during inference without retraining the model?

Answer: Web-augmented retrieval (search tool integration)

Web-augmented retrieval integrates a live search tool that the model can call during inference, grounding responses in current information without retraining.