V

VAE (Variational Autoencoder)

• Category: Models & inference
• What people say: A probabilistic generative autoencoder.
• What it actually means: A latent-variable model trained with a reconstruction objective and a regularization term that keeps an approximate posterior close to a chosen prior. The reparameterization estimator allows gradients through stochastic latent sampling.
• Common confusion: A VAE does not force every latent distribution to one fixed Gaussian; the exact prior and approximate posterior are modeling choices.
• Sources: Auto-Encoding Variational Bayes
• Related terms: Latent Space, Encoder, Decoder, Diffusion Model

Vector Database

• Category: Retrieval & generation
• What people say: A database optimized for vector similarity search.
• What it actually means: A storage and indexing system that supports nearest-neighbor queries over vector representations, often with metadata filtering, persistence, and approximate indexes.
• Common confusion: A vector database stores and searches vectors. It does not create high-quality embeddings or guarantee relevant retrieval.
• Related terms: Embedding, Semantic Search, Hybrid Retrieval

Verification Gate

• Category: Evaluation & safety
• What it actually means: A control point that blocks progress until defined evidence satisfies a correctness or quality criterion.
• Why it matters: It converts a model's claim of completion into an evidence-backed decision.
• In practice: Prevent a coding task from completing until the patch applies, scoped tests pass, forbidden files remain unchanged, and required artifacts exist.
• Common confusion: Verification checks whether evidence meets criteria. Approval grants authority to proceed, even when the evidence is already known.
• Learn it: Verification Gates
• Related terms: Approval Gate, Regression Test, Scope Contract, Structured Output

Vision-Language Model (VLM)

• Category: Multimodal systems
• What it actually means: A model that learns relationships between, or jointly processes, visual and language representations for tasks such as retrieval, description, question answering, or grounded generation.
• Why it matters: VLM performance depends on the visual encoder, language component, connection mechanism, training data, and resolution policy rather than one generic capability label.
• In practice: Evaluate text-only and vision-only controls, vary image resolution and layout, require evidence localization where possible, and report failures by visual skill and language.
• Common confusion: Accepting an image does not prove the model uses it correctly, and a VLM is not necessarily able to generate images.
• Learn it: Vision-Language Models
• Related terms: Multimodal Model, Vision Transformer (ViT), Cross-Attention, Visual Grounding
• Sources: CLIP; Flamingo

Vision Transformer (ViT)

• Category: Multimodal systems
• What it actually means: A vision architecture that represents an image as a sequence of patch embeddings with position information and processes that sequence with transformer encoder blocks.
• Why it matters: It provides a sequence-model interface for visual data, but performance and compute depend on patch size, resolution, pretraining, and inductive biases.
• In practice: Keep patching and normalization consistent with training, account for position-embedding behavior at new resolutions, and compare against a suitable visual baseline on the target dataset.
• Common confusion: ViT is an architecture family, not every transformer that accepts images, and its patches are not inherently semantic objects.
• Learn it: Vision Transformers
• Related terms: Transformer, Patch Embedding, Self-Attention, Encoder
• Sources: An Image is Worth 16x16 Words

Visual Grounding

• Category: Multimodal systems
• What it actually means: Connecting a language expression to spatial evidence in an image or video, such as a region, object, mask, or tracked entity.
• Why it matters: A fluent visual answer can be unsupported, while grounding makes the claimed referent inspectable and enables region-level evaluation.
• In practice: Require a box, mask, or temporal segment with the answer, test ambiguous and absent referents, and score localization separately from language correctness.
• Common confusion: Visual grounding identifies where the referenced evidence is. General image captioning can describe a scene without localizing each claim.
• Learn it: Cross-Attention Fusion
• Related terms: Grounding, Vision-Language Model (VLM), Attention, Evaluation (Eval)
• Sources: MDETR

Vocabulary

• Category: Data & representations
• What it actually means: The finite mapping between token identifiers and the units a tokenizer can emit, including ordinary, byte-level, and special control tokens.
• Why it matters: Vocabulary design affects sequence length, multilingual coverage, code representation, embedding size, and compatibility between tokenizers and model weights.
• In practice: Version the vocabulary and special-token assignments with the model, test encode-decode round trips, and never substitute a tokenizer with merely similar token names.
• Common confusion: A model vocabulary is not a dictionary of human words; many entries are fragments, bytes, whitespace patterns, or control symbols.
• Related terms: Tokenization, Byte Pair Encoding (BPE), Token, Embedding
• Sources: Neural Machine Translation of Rare Words with Subword Units