Glossary

A ranking metric that measures how well a system places relevant results near the top of a ranked list. It accumulates relevance gains across positions, applies a logarithmic discount to results appearing lower in the ranking, and normalizes the score against an ideal ordering, hence normalized discounted cumulative gain. Unlike binary metrics such as recall, NDCG supports graded relevance where some results are more relevant than others. Scores range from 0 to 1, with 1 representing a perfect ranking, making it the standard metric for retrieval benchmarks like MTEB.

Two pieces of content that should be treated as dissimilar during training. For example, a question is paired with an unrelated passage. The model learns to produce embeddings that are far apart for negative pairs.

The process of selecting negative examples for contrastive training. The quality of negatives significantly affects model performance. Random negatives are easy for the model and provide little learning signal; carefully selected hard negatives drive meaningful improvement.

A field of artificial intelligence concerned with enabling machines to understand, process, and generate human language. NLP underpins most tasks in text-based AI systems, from tokenization and parsing to semantic understanding and generation. Modern NLP is dominated by transformer-based models, which have largely replaced earlier rule-based and statistical approaches.

Scaling a vector so its length equals 1, producing a unit vector. After normalization, cosine similarity and dot product become equivalent, which simplifies computation. Most embedding models output normalized vectors by default.