RAG Systems Glossary
A comprehensive glossary of terms used in RAG (Retrieval-Augmented Generation) systems.
A
ANN (Approximate Nearest Neighbor): Algorithms that find approximately nearest neighbors faster than exact search, trading some accuracy for speed.
Answer Relevance: A metric measuring how well the generated answer addresses the original question.
B
BGE (Beijing Academy General Embedding): Open-source embedding models from BAAI known for high quality.
BM25: A keyword-based ranking algorithm used in traditional information retrieval systems.
C
Chunk: A segment of text created by splitting larger documents for indexing and retrieval.
Chunk Overlap: Shared text between consecutive chunks to preserve context at boundaries.
Chroma: An open-source, lightweight vector database popular for development.
Context Precision: RAGAS metric measuring relevance of retrieved documents.
Context Recall: RAGAS metric measuring if all necessary information was retrieved.
Cosine Similarity: Similarity metric measuring the angle between two vectors (range: -1 to 1).
Cross-Encoder: A model that scores query-document pairs together for reranking.
D
Dense Retrieval: Retrieval using dense vector representations (embeddings) rather than sparse keywords.
Document Loader: Component that extracts text from various file formats (PDF, DOCX, etc.).
E
Embedding: Dense vector representation of text that captures semantic meaning.
Euclidean Distance: Distance metric measuring straight-line distance between vectors.
F
FAISS (Facebook AI Similarity Search): High-performance library for similarity search and clustering of dense vectors.
Faithfulness: Metric measuring if the generated answer is grounded in retrieved context (no hallucinations).
Fine-tuning: Training a model on specific data to change its behavior or knowledge.
G
Generation: The phase where an LLM produces a response using retrieved context.
GPT (Generative Pre-trained Transformer): Family of large language models from OpenAI.
H
Hallucination: When an LLM generates false or unsupported information confidently.
HNSW (Hierarchical Navigable Small World): Popular graph-based algorithm for fast approximate nearest neighbor search.
HyDE (Hypothetical Document Embeddings): Technique of generating a hypothetical answer to improve retrieval.
Hybrid Search: Combining keyword (BM25) and semantic (vector) search.
I
Indexing: The offline process of preparing documents for retrieval (load, chunk, embed, store).
Inner Product (Dot Product): Similarity metric; for normalized vectors, equivalent to cosine similarity.
IVF (Inverted File Index): Vector search algorithm that clusters vectors for faster search.
K
Knowledge Cutoff: The date after which an LLM has no training data knowledge.
L
LangChain: Popular framework for building LLM applications including RAG pipelines.
LangSmith: Debugging and monitoring platform for LangChain applications.
LCEL (LangChain Expression Language): Declarative syntax for composing LangChain components.
LlamaIndex: Framework focused on data indexing and querying for LLM applications.
LLM (Large Language Model): Neural network models trained on large text corpora (GPT-4, Claude, Llama, etc.).
M
Metadata: Additional information stored with documents (source, date, author, etc.).
Milvus: Open-source, distributed vector database for large-scale deployments.
MMR (Maximum Marginal Relevance): Retrieval method balancing relevance with diversity.
MRR (Mean Reciprocal Rank): Metric measuring how high the first relevant document ranks.
MTEB (Massive Text Embedding Benchmark): Standard benchmark for comparing embedding models.
Multi-Query Retrieval: Generating multiple query variations to improve retrieval coverage.
Multi-Tenancy: Serving multiple organizations from a single platform with data isolation.
N
Namespace: Partition in a vector database for isolating data (e.g., per tenant).
O
OCR (Optical Character Recognition): Converting images of text into machine-readable text.
P
pgvector: PostgreSQL extension enabling vector similarity search.
Pinecone: Fully-managed cloud vector database.
Precision@K: Percentage of retrieved documents (top K) that are relevant.
Prompt: Text input to an LLM, often including instructions and context.
Q
Qdrant: High-performance vector database written in Rust.
Query Expansion: Adding related terms and synonyms to a query.
Query Rewriting: Using an LLM to rephrase queries for better retrieval.
R
RAG (Retrieval-Augmented Generation): Technique combining document retrieval with LLM generation.
RAGAS: Framework for automated RAG pipeline evaluation.
Recall@K: Percentage of all relevant documents that appear in top K results.
Reciprocal Rank Fusion (RRF): Method to combine rankings from multiple retrievers.
Reranking: Re-scoring initial retrieval results with a more powerful model.
Retrieval: The online phase of finding relevant documents for a query.
Retriever: Component that fetches relevant documents from a knowledge base.
S
Semantic Caching: Caching results for semantically similar queries.
Semantic Search: Finding documents based on meaning rather than exact keywords.
Self-Query: Using an LLM to extract both semantic query and metadata filters from input.
Sentence Transformers: Library and models for generating sentence embeddings.
Sharding: Distributing data across multiple database nodes.
T
Text Splitter: Component that divides documents into smaller chunks.
Token: Smallest unit of text processed by LLMs (roughly 4 characters in English).
Top-K: The K most similar/relevant results to return.
V
Vector: Numerical array representing text in embedding space.
Vector Database: Database optimized for storing and searching high-dimensional vectors.
Vector Store: Component that stores embeddings and enables similarity search.
W
Weaviate: Open-source vector database with built-in vectorization.
Quick Reference
| Term | Category | Definition |
|---|---|---|
| RAG | Core | Retrieval + Generation |
| Embedding | Core | Vector representation |
| Chunk | Processing | Text segment |
| HNSW | Algorithm | ANN search |
| RAGAS | Evaluation | RAG assessment |
| Hallucination | Problem | False output |
| Top-K | Retrieval | Number of results |