KI-Glossar

Vector Knowledge Base

Specialized storage optimized for embeddings, enabling fast, large-scale semantic similarity searches using indexes like HNSW or IVF.

HyDE (Hypothetical Document Embeddings)

An advanced technique where the LLM first generates an ideal hypothetical document, then uses its embedding to guide the search toward truly relevant documents.

Cross-Encoder Reranking

A relevance evaluation method where the model processes the query and candidate document simultaneously, unlike the standard bi-encoder approach.

Query Decomposition

A strategy where a complex query is automatically broken down into simpler sub-queries to improve the precision of multi-faceted information retrieval.

Synchronous vs Asynchronous RAG

The distinction between the synchronous approach (search and generation in the same call) and the asynchronous approach (pre-indexing and real-time retrieval) based on latency constraints.

Prompt Shaping

The art of optimizing the RAG prompt structure, including the placement of retrieved contexts, formatting instructions, and citation constraints to maximize response quality.

Multi-hop RAG

An advanced architecture where the model performs multiple retrieval-generation cycles, using intermediate responses to refine and deepen the information search.

Dynamic Knowledge Base

A RAG system where the document base is continuously updated in real-time, enabling always-current answers without requiring model retraining.

Domain-agnostic RAG

An approach where the retrieval system is designed to work effectively in any domain without specific adaptation, thanks to generalized embeddings and retrieval strategies.

Source citation

Hierarchical RAG

A multi-level architecture where retrieval first occurs on summaries or metadata, then on relevant detailed documents to optimize speed and relevance.

Context fusion

The process of intelligently integrating multiple retrieved documents into a coherent prompt, avoiding redundancies and maximizing the complementarity of information.

RAG with memory

An extension of standard RAG where the system maintains a memory of previous interactions to contextualize future retrievals and ensure conversational coherence.