The RAG stack — every term explained
Vector databases, embeddings, chunking, retrieval, reranking, evaluation. Every term you need to build, deploy, and debug a retrieval-augmented generation system.
Retrieval-Augmented Generation (RAG) is the most-deployed pattern for building LLM applications in 2026. Instead of relying only on what the model learned during training, RAG gives the model access to your documents at inference time — reducing hallucinations, enabling up-to-date answers, and grounding responses in cited sources.
The 70 terms in this category are grouped into four sub-topics: how you prepare data (chunking, embedding), how you store and retrieve it (vector databases, indexing), how you improve retrieval quality (reranking, hybrid search), and advanced production patterns (agentic RAG, multi-modal RAG, evaluation). For the "why" behind these terms, see our How RAG Works tutorial.
The most important terms in Vector Databases & RAG
Start here if you're new. These entries explain the foundational vocabulary in depth.
Retrieval-Augmented Generation (RAG)
The pattern of giving an LLM external documents at inference time by retrieving relevant chunks and injecting them into the prompt. Reduces hallucinations and enables up-to-date answers.
Read the full entry →FoundationsEmbedding
A dense vector representation of text, image, or other data. Semantically similar items produce similar vectors, enabling similarity search over meaning rather than keywords.
Read the full entry →StorageVector Database
A database optimized for storing and searching high-dimensional vectors. Pinecone, Weaviate, Qdrant, Chroma, and pgvector are common choices.
Read the full entry →PreparationChunking
Splitting documents into smaller passages that fit in the context window and embed meaningfully. Chunk size, overlap, and strategy dramatically affect RAG quality.
Read the full entry →ImprovementReranker
A second-stage model that re-scores retrieved candidates for relevance. Typically a cross-encoder that reads query + document together for better quality.
Read the full entry →RetrievalSemantic Search
Searching by meaning rather than keywords. The core RAG retrieval technique — find documents that mean what the query asks, not just contain the same words.
Read the full entry →All Vector Databases & RAG terms, organized
Grouped into sub-topics so you can find neighbors and prerequisites, not just alphabetical entries.
Data preparation
Storage & retrieval
Retrieval quality
Advanced RAG patterns
Related term categories
These categories connect naturally to Vector Databases & RAG — many terms cross-reference between them.
Go deeper on Vector Databases & RAG
📖 Related concept tutorials
Long-form guides that walk through how these concepts actually work.
- How RAG Works (End-to-End)
Complete walkthrough with code.
- How Embeddings Work
The math and intuition behind dense vectors.
- Choosing a Vector Database
Trade-offs across providers.
🎯 Related models & tools
Real products and models where you'll encounter these terms.
- Every LLM
RAG works with any LLM as generator.
- RAG tools
LangChain, LlamaIndex, and framework tools.
- RAG vs Fine-tuning
When to reach for each.
Questions about Vector Databases & RAG
No, but the split is shifting. Very long context windows reduce the need for RAG on small document collections. For large corpora, current information, or citation requirements, RAG is still essential. See our RAG vs Fine-tuning comparison.
Usually chunking. Bad chunks (too big, too small, split mid-idea) tank retrieval quality regardless of how good your embedding model is. Second-biggest is not using a reranker.
Depends on scale. Under ~100K documents, pgvector on PostgreSQL is usually fine. Above that, dedicated vector databases (Pinecone, Weaviate, Qdrant, Milvus) have advantages in query speed, updates, and metadata filtering.
Embedding is the technique (mapping text to vectors). Semantic search is the application (using those vectors to find similar meanings). One enables the other.
OpenAI's text-embedding-3-large is a strong default. For self-hosting, sentence-transformers models or BGE. For latest options, see the MTEB leaderboard. Embedding model choice is less impactful than chunking strategy in most systems.