Reference
LLM glossary
Short, precise definitions for the vocabulary used across LLM foundations, application engineering, and production systems.
- Agent
- A system that lets a model select actions within explicit tools, state, policy, and stop conditions.
- Attention
- A learned operation that mixes value vectors according to query-key compatibility scores.
- Chunk
- A bounded unit of source content indexed and retrieved as evidence.
- Context window
- The maximum token sequence a model can process in one request or generation step.
- Decoding
- The policy used to select output tokens from model scores.
- Embedding
- A learned vector representation used inside models or for semantic comparison and retrieval.
- Fine-tuning
- Additional training that adapts model parameters or adapters to a target behavior.
- Grounding
- Constraining an answer to identifiable evidence, tools, or verified state.
- Hallucination
- A fluent output that is unsupported, false, or inconsistent with required evidence.
- Inference
- Running a trained model to produce scores, embeddings, or generated output.
- KV cache
- Stored attention keys and values reused during autoregressive decoding.
- LoRA
- A parameter-efficient method that learns low-rank weight updates while freezing base weights.
- MCP
- Model Context Protocol, a standard for AI clients to discover and use external tools and resources.
- Prompt
- The instructions, context, examples, and data supplied to a model.
- RAG
- Retrieval-augmented generation, which supplies selected external evidence at answer time.
- Reranker
- A model or rule that reorders retrieved candidates using richer relevance evidence.
- Structured output
- A model response constrained to a machine-checkable schema.
- Temperature
- A decoding control that changes how concentrated the token probability distribution is.
- Token
- A discrete vocabulary unit represented by an integer ID.
- Tokenizer
- The versioned algorithm and vocabulary that convert text to token IDs and back.
- Top-p
- Nucleus sampling that considers the smallest high-probability token set reaching cumulative probability p.
- Tool call
- A structured request from a model-driven application to an external capability.
- Transformer
- A neural architecture built from attention, feed-forward layers, residual paths, and normalization.