Technical Glossary

Neural network architecture relying entirely on self-attention mechanisms to compute contextual representations in parallel.

Mathematical operation computing context weights by comparing Query, Key, and Value vector dot-products.

Sub-word atomic chunk created by tokenizers (e.g. BPE) mapping text strings to numerical vocabulary IDs.

Dense numerical vector representing semantic meaning of text tokens in a continuous high-dimensional vector space.

Architecture pattern augmenting LLM prompts with relevant context retrieved dynamically from external vector databases.

Maximum token capacity an LLM can digest in a single forward pass, spanning system prompts, history, and responses.

Hyperparameter scaling logit probabilities before softmax to tune generation randomness and entropy.

Sampling strategy filtering candidate tokens to the smallest set whose cumulative probability reaches threshold P.

Process of updating LLM weights on specialized task datasets to adjust domain behavior or response formats.

Parameter-efficient fine-tuning technique freezing base weights and inserting trainable low-rank decomposition matrices.

Alignment technique tuning LLM outputs using reward models trained on human pairwise preference rankings.

Database engine specialized for indexing and performing fast nearest-neighbor similarity searches over vector embeddings.

Mechanism enabling LLMs to emit structured JSON payload calls invoking external APIs, scripts, or database queries.