Glossary

Every term used anywhere in this guide, defined in one place — jump to the full explanation via the linked page

Why

A guide this dense inevitably uses a term on page 12 that was only briefly defined on page 3. This page exists as the single lookup point — every acronym and piece of jargon used anywhere in the RAG guide, defined in one or two sentences, with a link to the page that explains it in full depth.

Explanation
A – C
  • Agentic RAG — an LLM acting as an agent that decides whether and how many times to call a retriever, reasoning between calls. See Advanced Architectures.
  • ANN (Approximate Nearest Neighbor) — a family of search algorithms (HNSW, IVF, LSH) that trade a small amount of accuracy for a large speedup over brute-force vector comparison. See Vector Databases.
  • Answer Relevance — a RAGAS metric measuring whether a generated answer actually addresses the question asked. See Evaluation.
  • Bi-encoder — an embedding approach that encodes the query and each document separately, enabling fast, pre-computable similarity search. See Embeddings.
  • BM25 — a statistical keyword-ranking formula (a refinement of TF-IDF) used for lexical/keyword search. See Search Types.
  • Centroid — the average vector of a cluster, used by IVF and k-means to group similar vectors. See Vector Databases.
  • Chain strategies (Stuff / Map-Reduce / Refine / Map-Rerank) — four ways to combine multiple retrieved chunks into an LLM call. See Augmentation & Prompting.
  • Chunk / Chunking — a small piece of a document produced by a text splitter, and the process of producing it. See Text Splitters.
  • Chunk overlap — characters or tokens repeated between consecutive chunks, so a cut sentence isn't fully orphaned. See Text Splitters.
  • ColBERT — a "late interaction" retrieval model that keeps one vector per token, recovering much of a cross-encoder's precision at a fraction of the cost. See Advanced Retrieval.
  • Context poisoning — a security risk where untrusted retrieved text contains instructions the model may follow as if they came from the system. See Generation Failure Modes.
  • Context Precision / Context Recall — RAGAS metrics measuring, respectively, how much of what was retrieved is relevant, and whether everything relevant was retrieved. See Evaluation.
  • Context window — the maximum number of tokens an LLM can accept in a single request. See Context Window Limits.
  • Contextual compression — trimming each retrieved chunk down to only its relevant sentences before it reaches the prompt. See Augmentation & Prompting.
  • Contextual retrieval — prepending an LLM-generated summary of a chunk's place in its source document before embedding it. See Text Splitters.
  • Corrective RAG (CRAG) — grading retrieved chunks for relevance after retrieval and falling back to a broader search if they fail. See Advanced Architectures.
  • Cosine similarity — the angle between two vectors, the most common way to measure how close two embeddings are in meaning. See Embeddings.
  • Cross-encoder — an embedding approach that reads the query and a document together, far more accurate but far slower than a bi-encoder. See Embeddings.
D – H
  • Dense embedding — a vector where every dimension holds a learned, typically non-zero value, as opposed to a sparse embedding. See Embeddings.
  • Document loader — the pipeline stage that reads a raw file and converts it into a Document object. See Document Loaders.
  • Document object — the standard representation of a chunk of content: page_content plus metadata. See Document Loaders.
  • Dot product — a similarity measure related to cosine similarity, factoring in vector magnitude. See Embeddings.
  • Embedding — a vector representation of text (or an image) capturing its meaning, enabling semantic comparison via math. See Embeddings.
  • Euclidean (L2) distance — straight-line distance between two vectors; more common for image embeddings than text. See Vector Databases.
  • Faithfulness — a RAGAS metric measuring whether a generated answer is factually consistent with its retrieved context. See Evaluation.
  • fetch_k — the number of candidates an MMR retriever pulls before diversifying down to the final k. See Advanced Retrieval.
  • Fusion-in-Decoder — encoding retrieved passages separately and fusing them only in the decoder, allowing conditioning on more passages than fit as raw text. See Advanced Architectures.
  • GraphRAG — retrieval by traversing a knowledge graph of entities and relationships built from the corpus, instead of or alongside vector search. See Advanced Architectures.
  • Grounding — instructing (and structuring a prompt to ensure) an LLM answers from provided context rather than from memory. See Augmentation & Prompting.
  • Hallucination — an LLM generating fluent, confident, but factually incorrect or fabricated content. See What Is RAG?.
  • HNSW (Hierarchical Navigable Small World) — a graph-based ANN index, the default in most production vector databases. See Vector Databases.
  • Hybrid search — combining semantic (dense) and keyword (BM25/sparse) search, typically merged with Reciprocal Rank Fusion. See Search Types.
  • HyDE (Hypothetical Document Embeddings) — embedding an LLM-generated hypothetical answer instead of the raw query, to close a question/answer phrasing gap. See Advanced Retrieval.
I – M
  • Incremental indexing — updating or deleting only the vectors belonging to a changed document, instead of rebuilding the whole index. See Production Concerns.
  • Indexing phase — the offline, one-time-or-periodic half of the RAG pipeline: load, split, embed, store. See RAG Pipeline Overview.
  • IVF (Inverted File Index) — an ANN index that clusters vectors with k-means and searches only the nearest cluster(s). See Vector Databases.
  • k (top-k) — the number of chunks a retriever returns per query. See Retrieval Basics.
  • k-means clustering — an algorithm that groups vectors into k clusters around iteratively-recomputed centroids. See Vector Databases.
  • lambda_mult — the MMR parameter (0 to 1) controlling the relevance/diversity trade-off. See Advanced Retrieval.
  • LLM-as-judge — using a separate LLM to score another model's output against a rubric, as RAGAS's Faithfulness and Answer Relevance metrics do. See Evaluation.
  • "Lost in the middle" — the documented effect where models under-weight information buried in the middle of a long prompt. See Context Window Limits.
  • LSH (Locality-Sensitive Hashing) — an older ANN approach hashing similar vectors into shared buckets; largely superseded by HNSW for text. See Vector Databases.
  • Matryoshka embedding — an embedding trained so it can be truncated to fewer dimensions while still working reasonably well. See Embeddings.
  • Metadata — information about a chunk's origin (source, page, tags) that travels with it through the whole pipeline. See Document Loaders.
  • Metadata filtering — restricting a vector search to records matching specific metadata, before or after the similarity ranking. See Vector Databases.
  • MMR (Maximal Marginal Relevance) — a retrieval strategy balancing relevance against diversity among results. See Advanced Retrieval.
  • MQR (Multi-Query Retriever) — generating several rephrasings of a query and merging their separate retrieval results. See Advanced Retrieval.
  • MRR (Mean Reciprocal Rank) — an IR metric rewarding a relevant result appearing near the top of a ranked list. See Evaluation.
  • MTEB (Massive Text Embedding Benchmark) — the standard benchmark for comparing embedding models across tasks including retrieval. See Embeddings.
  • Multi-hop RAG — answering a question that requires multiple sequential retrieval steps, each depending on the last. See Advanced Architectures.
N – R
  • NDCG (Normalized Discounted Cumulative Gain) — an IR metric giving partial credit for ranking quality, penalizing relevant results found further down a list. See Evaluation.
  • Parent-document retriever — retrieving on small, precise chunks but returning their larger parent chunk to the LLM for more context. See Advanced Retrieval.
  • PII redaction — detecting and removing personal or sensitive information from documents before or during ingestion. See Production Concerns.
  • PQ (Product Quantization) — compressing vectors into smaller codes to save memory, used alongside IVF/HNSW at large scale. See Vector Databases.
  • Precision@k / Recall@k — classic IR metrics measuring, respectively, how many of the top-k results are relevant, and how many of all relevant items were found in the top-k. See Evaluation.
  • Prompt injection — text crafted to be interpreted as an instruction rather than data once it enters a prompt, a risk when retrieved content is untrusted. See Generation Failure Modes.
  • Propositional chunking — rewriting a document into self-contained, atomic factual statements, each its own retrievable unit. See Text Splitters.
  • RAGAS — an evaluation framework scoring RAG systems on Faithfulness, Answer Relevance, Context Precision, and Context Recall. See Evaluation.
  • RAPTOR — an indexing approach that builds a hierarchical tree of clustered, LLM-summarized chunks so retrieval can serve both fine-grained and gist-level questions. See Advanced Architectures.
  • ReAct (Reason + Act) — an agent pattern alternating between reasoning steps and tool-call actions, commonly used to implement Agentic RAG. See Advanced Architectures.
  • Reciprocal Rank Fusion (RRF) — a method for merging multiple ranked result lists (e.g., BM25 and dense) by rank position rather than raw score. See Search Types.
  • Recursive character text splitter — a splitter that tries progressively smaller natural-language separators (paragraph, line, sentence, word) before falling back to raw characters. See Text Splitters.
  • Reranking — re-scoring a retrieved shortlist with a slower, more accurate cross-encoder before generation. See Advanced Retrieval and the ranking-vs-reranking distinction in Retrieval Basics.
  • Retriever — a runnable component that takes a query and returns the most relevant chunks, regardless of the strategy underneath. See Retrieval Basics.
S – V
  • Self-query retriever — using an LLM to parse a natural-language query into a semantic search string plus structured metadata filters. See Advanced Retrieval.
  • Self-RAG — a model trained to generate reflection tokens that decide when to retrieve and critique its own retrieved passages and output. See Advanced Architectures.
  • Semantic chunking — splitting text at points where sentence-embedding similarity drops sharply, i.e. where the topic shifts. See Text Splitters.
  • Semantic search — retrieval based on embedding similarity (meaning) rather than exact term matching. See Search Types.
  • Sharding / Replication — splitting a vector index across multiple nodes, and duplicating it, to scale storage, throughput, and fault tolerance. See Vector Databases.
  • Sparse embedding / SPLADE — a vector representation where most values are zero and non-zero values roughly correspond to specific terms, blending keyword precision with learned relevance. See Embeddings.
  • Structured output — requesting a schema-constrained (e.g., JSON) response from generation instead of free text, so answer, citations, and confidence are machine-readable. See Augmentation & Prompting.
  • Token — the unit an LLM or embedding model actually counts length in; roughly 4 characters per token in English. See Text Splitters.
  • Vanilla retriever — the simplest retriever: embed the query, run cosine similarity search, return the top-k. See Retrieval Basics.
  • Vector database — a database purpose-built to store embeddings and their metadata and answer nearest-neighbor queries fast. See Vector Databases.
← Previous Practice interview questions →