BlogGlossary

Glossário de IA em linguagem simples

Definições curtas das palavras que aparecem em ferramentas, cursos e anúncios de emprego.

A

A/B test: Comparing two versions by measuring outcomes.
Agent: Software que pode assumir um objetivo e usar ferramentas por passos.
Alignment: Work to make a model follow intended behavior, including refusals. Not a truth guarantee.
Annotation: Labels humans add to data so models can learn or so teams can evaluate answers.
API: Uma forma de um programa pedir a outro para fazer algo.
Artificial intelligence: Software that drafts, classifies, or spots patterns in ways we associate with judgment.
Attention: A transformer mechanism that lets the model weigh which tokens matter for the next prediction.
Automation: A saved series of steps that runs with less manual work.
Autonomous agent: An agent allowed to take multiple tool actions with less step-by-step human steering.

B

Batch inference: Running many prompts together for efficiency.
Beam search: A decoding strategy that keeps several candidate sequences.
Benchmark: A standard test used to compare models. High scores are not the same as fit for your files.
Bias: A systematic skew in data or output, including unfair patterns from training data.

C

Caching: Reusing earlier results to save time or cost.
Chain of thought: Visible intermediate reasoning steps. Helpful for hard problems; still not a source check.
Chatbot: An interface where you type and a model replies in turns.
Checkpoint: A saved snapshot of model weights during or after training.
Chunking: Splitting documents into pieces for retrieval.
Citation: A link or reference back to a source. Useful only if you open it.
Classification: Assigning an item to a category, such as spam or not spam.
Claude: Anthropic’s family of assistants for chat and, on some plans, document tools.
Completion: The model’s continuation of a prompt.
Computer use: When an agent is allowed to operate a screen or browser like a user.
Constitutional AI: A training approach that uses written principles to shape behavior.
Content filter: A classifier that flags disallowed material.
Context stuffing: Packing too much into the window so important bits get lost.
Context window: Quanto texto um modelo consegue considerar de uma vez.
Contrastive learning: Learning by pulling similar items together and pushing others apart.
Copilot: A name for assistants inside other products, often Microsoft’s in Microsoft 365.
Corpus: A large collection of texts used for training or evaluation.
Cosine similarity: A score for how close two embeddings are.

D

Data leakage: When evaluation data accidentally appears in training.
Dataset: A collection of examples used to train or test a model.
Decoding: Turning model scores into actual tokens.
Deterministic: Producing the same output for the same input when configured to.
Diffusion model: An image model that turns noise into a picture guided by a prompt.
Distillation: Training a smaller model to imitate a larger one.

E

Embedding: A list of numbers representing text or an image for similarity search.
Embedding space: The map where similar meanings sit near each other.
Endpoint: A URL where an API receives requests.
Ensemble: Combining several models or prompts to improve robustness.
Entropy: A measure of uncertainty in predictions.
Epoch: One full pass through a training dataset.
Evaluation: A repeatable way to measure whether a change helped.

F

Feature store: A system that serves prepared signals to models.
Few-shot: Providing a few examples in the prompt.
Fine-tuning: Further training on a narrower set of examples.
Foundation model: A large model trained broadly, then adapted into products.
Function calling: Letting a model request a structured tool call the app executes.

G

Generative AI: Models that create text, images, audio, or code from a prompt.
GPU: Hardware often used to train and run large models quickly.
Gradient: A signal used during training to update weights.
Ground truth: The accepted correct answer used for evaluation.
Grounding: Tying answers to specific sources such as uploaded files.
Guardrails: Product rules that filter or reshape unsafe or off-policy outputs.

H

Hallucination: Uma resposta fluente falsa, sem suporte ou inventada.
Heuristic: A practical rule that is not guaranteed optimal.
Human in the loop: A person reviews or approves output before it counts.
Hybrid search: Combining keyword and semantic search.
Hyperparameter: A training setting chosen by people, such as learning rate.

I

Inference: Using a trained model to produce an answer.
Inference cost: What you pay in money or compute per answer.
Instruction tuning: Training that teaches a model to follow instructions better.

J

Jailbreak: An attempt to bypass model safety rules.
JSON mode: Forcing output into JSON so software can parse it.

K

Knowledge base: Documents an app can retrieve from when answering.
Knowledge cutoff: The date after which training data may not include newer events.
KV cache: A speed optimization that stores intermediate attention values.

L

Latency: How long you wait for a reply.
Latency budget: How long a product can wait before users feel delay.
LLM: Large language model — a model trained on lots of text.
Logprob: The log probability of a token under the model.
LoRA: A lightweight fine-tuning method that adds small adapter weights.

M

Machine learning: Métodos que aprendem padrões a partir de exemplos.
Memory: Product features that store facts about a user across chats.
Modality: A kind of input or output such as text, image, or audio.
Model card: A document describing intended use, limits, and training notes.
Model routing: Sending a request to different models based on the task.
MoE: Mixture of experts — routing tokens to specialized subnetworks.
Multimodal: Accepting or producing more than one kind of data.

N

N-gram: A short sequence of n tokens or words.
Natural language processing: Computing with human language.
Neural network: A layered model architecture used in deep learning.
Noise: Randomness or irrelevant variation in data or generation.

O

OCR: Optical character recognition — reading text from images.
On-device model: A model that runs on a phone or laptop instead of a remote server.
Ontology: A structured map of concepts and relationships.
Open-source model: A model shared under a license you can read; check commercial terms.
Orchestration: Code that coordinates prompts, tools, and checks.
Overfitting: Learning training examples so narrowly that new cases fail.

P

Parameter: A number inside a model set during training.
Perplexity: In evaluation, how surprised a model is by text; also a product name.
Persona: A standing style or role you ask the assistant to take.
PII: Personally identifiable information that needs careful handling.
Plugin: An extension that lets an assistant call an external service.
Post-processing: Rules applied after generation, such as formatting.
Prefill: The phase of inference that reads the prompt before decoding.
Pretraining: The first large training pass before specialization.
Privacy budget: Limits on how much private signal a system may expose.
Prompt: As instruções e o contexto que dá a um modelo.
Prompt engineering: Writing and revising prompts so outputs are usable.
Prompt injection: Malicious text that tries to override instructions.

Q

Quantization: Storing weights with fewer bits to run cheaper or faster.

R

RAG: Geração aumentada por recuperação: procura nos seus documentos e depois gera.
Rate limit: A cap on how many requests you can send.
Re-ranking: Reordering retrieved passages with a stronger model.
Reasoning model: A mode that spends extra steps before answering.
Red teaming: Deliberately probing a system for failures and harms.
Refusal: When a model declines a request under its policy.
Regression test: An evaluation that catches answers that used to work and now fail.
Reinforcement learning: Learning from rewards and penalties.
Retrieval: Finding relevant passages for a question.
RLHF: Reinforcement learning from human feedback.

S

Safety filter: A layer that blocks or rewrites disallowed content.
Sampling: Choosing the next token from a probability distribution.
Sandbox: An isolated environment for running tools safely.
Schema: A description of allowed fields in structured output.
Seed: A number that makes random sampling repeatable.
Self-consistency: Sampling several answers and picking the common one.
Semantic search: Search by meaning using embeddings, not only keywords.
Sentiment analysis: Estimating attitude in text, such as positive or negative.
Sequence: An ordered list of tokens the model processes.
Serving: Running models in production for live traffic.
Shadow mode: Logging what a new system would do without acting yet.
Speculative decoding: A speed trick using a small model to propose tokens.
Speech to text: Turning audio into written words.
Stop sequence: Text that tells the model to end a completion.
Streaming: Sending the answer token by token as it is generated.
Structured output: An answer forced into a shape such as a list or JSON.
Style transfer: Rewriting text or images to match a target style.
Supervised learning: Learning from labeled examples.
Synthetic data: Data generated by models or scripts for training or tests.
System prompt: Standing instructions that shape assistant behavior.

T

Telemetry: Product usage metrics used to improve systems.
Temperature: Uma definição que muda quão variado o texto pode ser. Não é um dial de precisão.
Text to speech: Turning written words into audio.
Token: Um pedaço de texto que o modelo conta, muitas vezes parte de uma palavra.
Tokenizer: The component that splits text into tokens.
Tool use: When a model can call calculators, search, or other apps.
Top-k: Sampling only from the k most likely next tokens.
Top-p: Sampling from the smallest set of tokens whose probabilities sum to p.
Trace: A record of prompts, tools, and outputs for debugging.
Training data: Examples a model learned from.
Transformer: The architecture behind most modern language models.
Truncation: Cutting text that does not fit the context window.

U

Unit test for prompts: Fixed examples that must keep passing after prompt changes.
Unsupervised learning: Learning structure without labels.

V

Vector database: A store optimized for embeddings and similarity search.
Vision model: A model that works with images or video frames.
Vision-language model: A model that jointly handles images and text.

W

Warm start: Starting from existing weights instead of random ones.
Watermarking: Signals meant to mark generated content, with limited reliability.
Web grounding: Fetching live pages before answering.
Weights: The saved parameters of a trained model.
Whisper: A common name for speech-to-text models in that family.
Workflow: Origem → ferramenta → verificação → destino de um trabalho.

X

XML tagging: Using tags in prompts to separate instructions from data.

Y

YOLOs: Informal term for shipping generated code without review — avoid it.

Z

Z-shot: Another way some people say zero-shot.
Zero-shot: Pedir uma tarefa só com instruções, sem exemplos.