AI & ML Tech Glossary

A repetition penalty is a decoding control that lowers the probability of reusing recently generated tokens or phrases, reducing looping and repetitive LLM outputs. It...

Rotary positional embedding encodes token positions by rotating transformer query and key vectors so attention depends on relative position. It’s widely used in modern LLMs...

RLHF aligns a model with human preferences by training a reward model from human judgments and then using reinforcement learning to optimize the model to...

RAG groundedness measures how well an LLM’s answer is supported by the retrieved context. It focuses on claim-level faithfulness and citation support, helping reduce hallucinations...

Reranking is a second-stage retrieval step that reorders initially retrieved RAG chunks using a stronger relevance model (often a cross-encoder) so the LLM receives higher-precision...

RAG query rewriting transforms a user question into better retrieval queries before searching a vector database or index. It resolves ambiguity, expands context, and can...

RAG hallucination mitigation refers to techniques that reduce unsupported statements in retrieval augmented generation systems by improving retrieval quality, forcing evidence based answers with citations,...

Retrieval-augmented generation (RAG) grounds an LLM’s output in external documents retrieved at query time, improving factuality, freshness, and traceability compared with using the model’s parameters...

RAG chunking is the step of splitting documents into retrievable pieces before indexing for retrieval-augmented generation. Chunk size, overlap, and split boundaries strongly impact retrieval...

Reward modeling trains a model to score AI outputs based on human preferences or policy goals. The reward score then guides alignment methods like RLHF,...

Retrieval evaluation (RAG evaluation) measures how well a retriever finds the right evidence for a query and how that affects the final RAG answer. It...