Glossary
Vibe coding glossary
The everyday vocabulary of vibe coders, A to Z. Each entry is one or two sentences with links to related terms and a reference where useful.
66 terms
A
3 terms- Agent
- An LLM that runs in a loop, picking tools, executing them, reading results, and deciding the next step until it returns a final answer or hits a budget.
- Agent loop
- The runtime that turns model output into tool calls, runs them, feeds results back, and stops when the model returns a non-tool response or the call budget is hit.
B
2 terms- Batch inference
- Sending many prompts in one request to amortise tokenisation and network cost. Anthropic and OpenAI both offer cheaper batch endpoints with relaxed latency SLAs.
- BERT
- A 2018 encoder-only transformer from Google. Reads text bidirectionally and outputs embeddings. Largely superseded by decoder-only LLMs but still common as a reranker.
C
6 terms- Caching (prompt cache)
- Marking part of the prompt as reusable so repeated calls within a TTL only pay for the new tokens. Big system prompts benefit the most.
- Chain of thought (CoT)
- Asking the model to write its reasoning before its final answer. Boosts accuracy on hard tasks; many frontier models do it implicitly now via 'thinking' modes.
- Completion
- The text the model generates in response to a prompt. Older API name (single-turn); modern APIs use 'message' for multi-turn chat.
- Constitutional AI
- Anthropic's training approach where a model critiques and rewrites its own outputs against a written set of principles, instead of relying on human-labelled comparisons.
- Context engineering
- The practice of choosing what goes into the model's context window — system prompt, retrieved docs, tool descriptions, prior turns — to maximise the right answer per token spent.
- Context window
- Maximum number of tokens the model can read in one call. 200k for Claude Sonnet 4.6, 1M+ for Gemini 2.5 Pro. Includes system prompt, history, retrieved docs, and the answer.
D
3 terms- Decoder-only
- The transformer architecture used by all modern chat LLMs. Generates text left-to-right, one token at a time. Encoder-decoder (T5, BART) is mostly retired for chat.
- Diffusion model
- Generative model class that turns random noise into images, audio, or video by iteratively denoising. Powers Stable Diffusion, FLUX, Sora-style video.
- Distillation
- Training a smaller 'student' model to imitate a larger 'teacher' model's outputs. Cheaper to serve at the cost of some quality. Haiku-class models are often distilled from Sonnet-class.
E
4 terms- Edge runtime
- JavaScript runtime that runs at CDN PoPs (Vercel Edge, Cloudflare Workers). Cold-start in milliseconds; subset of Node APIs; ideal for low-latency reads worldwide.
- Embedding
- A fixed-size vector that encodes the meaning of a piece of text or image. Cosine distance between embeddings measures semantic similarity. The basis of search, RAG, recommendations.
- Eval
- Automated test that scores a model or prompt on a fixed set of inputs. Catches regressions when you tweak the prompt, swap models, or update the system prompt.
- Evaluation
- Broader process of measuring AI quality: accuracy, helpfulness, harmlessness, latency, cost. Uses both deterministic checks and LLM-as-judge scorers.
F
2 terms- Fine-tuning
- Continuing to train a base model on your own labelled data to bias it toward your domain or style. Heavier than prompting, lighter than training from scratch.
- Function calling
- Telling the model the JSON schema of functions it can invoke. The model returns a structured tool_use block; your runtime executes it and feeds back the result.
G
1 terms- Generative
- AI that produces new content (text, image, audio, video) rather than only classifying or scoring existing content.
H
3 terms- Hallucination
- When the model writes plausible-sounding text that isn't grounded in the input or in fact. Mitigated with retrieval (RAG), citations, narrower prompts, and evals.
- HNSW
- Hierarchical Navigable Small World — the dominant ANN (approximate nearest neighbour) index for vector search. Used inside pgvector, Qdrant, Weaviate, Pinecone.
I
3 terms- In-context learning
- Showing the model examples inside the prompt to teach it a pattern, with no weight updates. The cheapest way to specialise a base model.
- Inference
- Running a trained model to get a prediction. Distinct from training. The cost line item that grows with usage.
- Instruction tuning
- Fine-tuning a base model on examples of (instruction, ideal answer). Turns a raw next-token predictor into something that follows orders.
J
1 terms- JSON mode / strict output
- Constraining the model to emit syntactically valid JSON or to match a JSON Schema. Anthropic does it via system prompt + jsonrepair fallback; OpenAI has a `response_format` flag.
K
1 terms- KV cache
- Per-request cache of attention keys/values that lets the model decode each new token in O(1) over the existing prompt. The reason re-running the same prompt is fast.
L
4 terms- Latent space
- The model's internal high-dimensional representation. Embeddings live here. Operations like vector arithmetic ('king − man + woman ≈ queen') happen here.
- LLM
- Large Language Model. A neural network trained on huge text corpora to predict the next token. Modern frontier examples: Claude, GPT-4, Gemini, Mistral, Llama.
- llms.txt
- Proposed convention: a `/llms.txt` file at the root of a site that lists the URLs and short descriptions an LLM should read for context. Like robots.txt for AI.
- LoRA
- Low-Rank Adaptation. Cheap fine-tuning that learns a small low-rank delta on top of frozen base weights. The dominant PEFT method for open-weight models.
M
4 terms- MCP
- Model Context Protocol. Open standard from Anthropic that lets LLM clients (Claude Desktop, Cursor) talk to external 'servers' for tools, resources, and prompts.
- Mixture of Experts (MoE)
- Architecture where each token routes through only a few specialised sub-networks, not the whole model. Lets you scale parameters without scaling per-token compute. Used in DeepSeek-V3, Mixtral.
- Model card
- Public document describing a model's training data, intended uses, limitations, evals, and safety mitigations. Industry standard since 2019.
- Multi-modal
- Model that accepts more than one input type — text + images, text + audio, sometimes video. GPT-4o, Gemini 2.5, and Claude 4.x all support image input.
O
2 terms- One-shot prompting
- Showing the model exactly one example of the desired input → output before asking for the real answer.
- OpenAI-compatible API
- Many providers (Together, Groq, Mistral, OpenRouter) expose an API that mirrors OpenAI's `/v1/chat/completions` shape so you can swap providers by changing the base URL.
P
5 terms- PEFT
- Parameter-Efficient Fine-Tuning. Umbrella for techniques (LoRA, prefix tuning, IA³) that train a tiny number of new parameters instead of all of them.
- pgvector
- Postgres extension that adds a `vector` column type plus IVF/HNSW indexes. Lets you do RAG inside the same DB as your business data — no separate vector store needed for ≤1M vectors.
- Prompt
- The text you send to the model. Splits into the system prompt (persistent instructions) and user prompt (the current turn).
- Prompt caching
- Same as caching: reuse the encoded representation of a long fixed prefix across calls. Anthropic supports it via `cache_control: ephemeral` blocks.
- Prompt injection
- Untrusted input that smuggles instructions to the model ('ignore previous instructions and ...'). The classic security failure mode for any LLM that reads external content.
Q
1 terms- Quantization
- Reducing the precision of model weights (fp16 → int8 → int4) so the model fits on smaller GPUs or laptops. Trade some quality for big speed and memory wins.
R
4 terms- RAG
- Retrieval-Augmented Generation. Embed your docs, store them in a vector DB, retrieve the top-k for each query, and stuff them into the model's context. The standard way to ground answers in your data.
- Reasoning model
- A model that spends extra inference compute on internal 'thinking' (visible or hidden) before answering. Examples: o1/o3 series, Claude 4 thinking mode, DeepSeek-R1.
- Reranker
- A second-stage model that re-orders candidates from a fast first-stage retriever. Common in search and RAG pipelines; Cohere Rerank and BGE rerankers are popular.
- RLHF
- Reinforcement Learning from Human Feedback. Train a reward model from human comparisons, then fine-tune the LLM to maximise the reward. Used in InstructGPT and many post-training stages.
S
6 terms- Sampling
- How the next token is picked from the probability distribution. Greedy = always max. Sampling with temperature / top-k / top-p adds controlled randomness.
- Semantic search
- Search by meaning rather than keyword overlap. Embed the query and the documents; rank by cosine similarity. Pairs well with classic BM25 in a hybrid setup.
- Server-Sent Events (SSE)
- One-way streaming protocol over HTTP. The transport every LLM 'streaming' API uses: server pushes tokens as they're decoded, client renders incrementally.
- Serverless
- Compute model where you don't manage VMs — the platform spins up an instance per request and shuts it down after. Vercel functions, AWS Lambda, Cloudflare Workers.
- SFT
- Supervised Fine-Tuning. The standard fine-tune: feed (prompt, ideal answer) pairs and minimise loss. Usually the first post-training step before RLHF or DPO.
- Structured output
- Constraining the model to return data that matches a schema (JSON, XML, function arguments). Removes the need for fragile regex parsing.
T
5 terms- Temperature
- Sampling knob: 0 = deterministic / always pick the highest-probability token, higher = more creative / more random. Most production code lives between 0 and 0.7.
- Tokenizer
- Splits text into tokens before the model sees it. Different families (BPE for GPT, SentencePiece for Llama, Tiktoken for OpenAI) split slightly differently — the same string can be 50 or 70 tokens.
- Tool use
- The LLM emits a structured request to call one of the developer-provided tools. The runtime executes it and returns the result. Foundation of every agent.
- Top-k / top-p (nucleus) sampling
- Two ways to truncate the next-token distribution. Top-k keeps the k highest-probability tokens; top-p keeps the smallest set whose cumulative probability ≥ p. Tame randomness without going fully greedy.
- Transformer
- Neural network architecture from the 2017 'Attention Is All You Need' paper. The base of every modern LLM. Self-attention lets each token attend to every other token in the context.
U
1 terms- UMAP
- Dimensionality reduction algorithm. Projects high-dim embeddings into 2-D so you can plot semantic clusters. Faster and often clearer than t-SNE.
V
2 terms- Vector database
- Store and search embeddings at scale. Popular options: Pinecone, Qdrant, Weaviate, Milvus, Chroma, plus pgvector inside Postgres. Powers RAG and semantic search.
- Voice agent
- End-to-end pipeline that listens (STT), reasons (LLM), and speaks (TTS) — often over WebRTC for sub-second latency. Built with platforms like Vapi, Retell, or Pipecat.
W
2 terms- WebRTC
- Peer-to-peer real-time audio/video standard built into browsers. The transport for low-latency voice agents and many Lovable-style co-edit demos.
- Weights
- The numeric parameters of a trained model. 'Open weights' means the file is downloadable and runnable; 'open source' adds the training code and data.
Z
1 terms- Zero-shot prompting
- Asking the model to do a task with no examples — just the instruction. Most modern frontier models are strong zero-shot.