AI & ML Tech Glossary

A safety classifier detects policy-violating or sensitive content in AI systems (inputs, outputs, or agent tool actions) and is used to block, route, redact, or...

Self-RAG is a RAG method where the model decides when to retrieve, scores retrieved evidence for relevance and support, and uses those signals to produce...

Structured prompting constrains a model’s input and output to a defined format, such as labeled sections or JSON. By specifying a schema and rules, it...

Safety alignment is the practice of shaping an AI model so its responses and actions follow human values and safety policies. It uses methods like...

A system prompt is a high-priority instruction given to an LLM that sets its role, rules, and safety boundaries for the entire conversation. It typically...

Synthetic data generation creates artificial datasets that mimic real data’s statistical properties without copying individual records. It is used to train and test AI models...

Semantic caching reuses LLM responses for new requests that are meaningfully similar, not just text-identical. It uses embeddings and vector similarity search to decide when...

Self-consistency decoding samples multiple reasoning outputs from an LLM and selects the most consistent final answer using voting or aggregation. It often improves accuracy on...

Structured outputs constrain an LLM to produce schema-valid, machine-readable results (often JSON), improving reliability for extraction, automation, and tool-using agents by reducing parsing errors.

Speculative decoding speeds up LLM inference by having a small draft model propose several next tokens and a larger target model verify them in batches....