KI-Glossar
Das vollständige Wörterbuch der Künstlichen Intelligenz
Dense Passage Retrieval (DPR)
Dense retrieval architecture specifically designed to extract relevant passages, using separate BERT encoders for questions and passages with contrastive learning.
Bi-Encoder Architecture
Retrieval architecture using two independent encoders for queries and documents, allowing pre-computed indexing of documents for large-scale search.
Cross-Encoder Architecture
Architecture where the query and document are concatenated and processed together by the same encoder, offering superior accuracy at the expense of search speed.
Hierarchical Navigable Small World (HNSW)
Vector indexing algorithm using multi-level graphs for approximate nearest neighbor search with an excellent trade-off between speed and accuracy.
Maximal Marginal Relevance (MMR)
Search result diversification algorithm balancing relevance and novelty, selecting documents that maximize relevance while minimizing semantic redundancy.
Hybrid Search
Approach combining dense and sparse retrieval to leverage the respective strengths of semantic search and keyword search, improving overall accuracy.
Sentence Transformers
BERT models specifically trained to generate high-quality sentence embeddings, optimized for semantic similarity tasks and clustering.
Vector Indexing
Process of organizing dense vectors in specialized data structures to accelerate similarity queries, essential for large-scale search.
Dense Retrieval Augmented Generation
Extension of RAG specifically using dense retrieval to provide relevant context to generation models, improving the coherence and factual accuracy of generated responses.
Late Interaction
Retrieval paradigm where interactions between query and document occur late in the process, after separate encoding, allowing a trade-off between precision and computational efficiency.