Every RAG term and abbreviation, expanded and explained in one place — from ANN and BM25 to nDCG, RRF, and reflection tokens

Glossary

How to use this page

Every term and abbreviation used anywhere in this section, expanded and explained in plain language. Organised A–Z. Where a concept has a full explanation elsewhere, the entry links to it.

Property	Value
Level	All levels — reference
Reading time	Reference; skim or search
Prerequisites	None
You will understand	Every piece of RAG vocabulary you are likely to encounter

A

Term	Expansion	Meaning
ACL	Access Control List	Metadata on a chunk recording who may see it. Must be applied as a retrieval-time filter, not after retrieval
Adaptive RAG	—	Routing each query to the cheapest sufficient path based on its complexity. See Architectures
Agentic RAG	—	Retrieval exposed as a tool an agent may call repeatedly on its own initiative. Powerful, slow, expensive
ANN	Approximate Nearest Neighbour	Search that finds almost the closest vectors, trading a little accuracy for an enormous speed gain. What makes vector databases fast
Answer Relevance	—	Evaluation metric: does the answer address the question actually asked?
Attribution marker	—	The `[1]`, `[2]` labels attached to chunks in the context so the model can cite them. Without them, citations get invented
Augmentation	—	The "A" in RAG: inserting retrieved text into the prompt

B

Term	Expansion	Meaning
Bi-encoder	—	A model that embeds query and document separately, so document vectors can be precomputed. Fast, less accurate. Contrast cross-encoder
BM25	Best Matching 25	The standard keyword-ranking algorithm. Scores documents on term frequency, adjusted for term rarity and document length. Excellent at exact and rare terms
Blue/green migration	—	Building a new index alongside the old one and cutting over, so rollback stays possible. The only safe way to change embedding models

C

Term	Expansion	Meaning
Chunk	—	One retrievable piece of a document. The unit that gets embedded, stored, searched, and inserted into the prompt
Chunking	—	Splitting documents into chunks. See Chunking Strategies
ColBERT	Contextualized Late Interaction over BERT	Late-interaction retrieval: keeps a vector per token rather than per chunk, matching at token level. More accurate than bi-encoders, much larger storage
Context Precision	—	Are the relevant chunks ranked above the irrelevant ones in the retrieved set?
Context Recall	—	Of the facts needed for the correct answer, how many are present in the retrieved context?
Context window	—	The maximum tokens a model can process in one call — prompt plus output
Contextual retrieval	—	Prepending a short generated summary of the surrounding document to each chunk before embedding, so the chunk is interpretable standalone
Corpus	—	The full collection of documents your system can retrieve from
Cosine similarity	—	Similarity between two vectors measured as the angle between them, ignoring magnitude. The standard measure for text embeddings. `1.0` identical, `0` unrelated
CRAG	Corrective RAG	Grades retrieved documents before generating, and falls back — often to web search — when the grade is poor
Cross-encoder	—	A model that reads query and document together and scores their relationship. Far more accurate than a bi-encoder, far too slow to run over a whole corpus. The standard reranker

D

Term	Expansion	Meaning
Dense retrieval	—	Search by embedding similarity — matching by meaning. Contrast sparse retrieval
Dimensions	—	The length of an embedding vector, e.g. 384, 768, 1536. More dimensions capture more nuance and cost more storage and compute
Distance metric	—	How vector closeness is computed: cosine, dot product, or Euclidean (L2)
Document ID	—	A stable identifier linking chunks back to their source document. Required to delete superseded chunks on update
Dot product	—	A similarity measure that accounts for both angle and magnitude. Equivalent to cosine when vectors are normalized

E

Term	Expansion	Meaning
efSearch	—	HNSW parameter controlling how much of the graph is explored at query time. Higher means better recall and higher latency
Embedding	—	A list of numbers representing the meaning of a piece of text. Similar meanings produce similar vectors. See Embeddings Explained
Embedding model	—	The model that converts text into vectors. Query and documents must use the same one
Extractive vs abstractive	—	Extractive answers quote source text; abstractive answers rephrase it. RAG is usually abstractive with extractive grounding

F

Term	Expansion	Meaning
FAISS	Facebook AI Similarity Search	A widely-used open-source library for fast vector similarity search. A library, not a database — no persistence or filtering of its own
Faithfulness	—	Evaluation metric: is every claim in the answer supported by the retrieved context? The primary anti-hallucination measure
Fine-tuning	—	Modifying a model's weights to change its behaviour. Teaches style and skill; RAG supplies facts. Usually complementary, not alternatives
Fixed-size chunking	—	Splitting by a fixed character or token count. Simple, and blind to meaning

G

Term	Expansion	Meaning
Golden dataset	—	The labelled evaluation set: questions, ground-truth answers, and the chunk IDs that should be retrieved
Graph RAG	—	Retrieval over a knowledge graph of entities and relationships, answering questions whose answers live between documents
Grounding	—	Constraining the model to answer only from retrieved context
Groundedness	—	See Faithfulness

H

Term	Expansion	Meaning
Hallucination	—	A fluent, confident, false statement. RAG reduces the rate and makes it detectable; it never eliminates it
Hit Rate@k	—	Did any relevant chunk appear in the top k? The coarsest retrieval metric — easily flattering
HNSW	Hierarchical Navigable Small World	The most common ANN index. A layered proximity graph navigated coarse-to-fine. Excellent recall and speed; memory-hungry
Hybrid search	—	Running dense and sparse search together and fusing the rankings. The correct default for production
HyDE	Hypothetical Document Embeddings	Have the model write a fake answer, embed that, and search with it — so answer-shaped text matches answer-shaped documents

I

Term	Expansion	Meaning
IDCG	Ideal Discounted Cumulative Gain	The DCG of a perfect ranking; the denominator that normalizes nDCG
Indexing pipeline	—	The offline pipeline: load, clean, chunk, embed, store. Runs when documents change
Indirect prompt injection	—	An attack where malicious instructions are planted inside a document your system indexes and later retrieves into the prompt
Ingestion	—	Getting source documents into the system: loading, parsing, cleaning
IVF	Inverted File Index	An ANN index that clusters vectors and searches only the nearest clusters. Lower memory than HNSW; needs a training step

K

Term	Expansion	Meaning
k	—	How many results a search returns. `top-k` is the k highest-scoring results
Knowledge cutoff	—	The date after which a model's training data ends. One of the core problems RAG solves

L

Term	Expansion	Meaning
L2 distance	Euclidean distance	Straight-line distance between two vectors. Less common than cosine for text
Late chunking	—	Embedding the whole document first, then pooling token embeddings into chunks — so each chunk vector retains document-wide context
Late interaction	—	Matching at token level at query time rather than comparing single pooled vectors. See ColBERT
LLM	Large Language Model	The model that writes the final answer. Supplies language ability; your documents supply the facts
LLM-as-judge	—	Using a model to score outputs on subjective criteria like faithfulness. Must be validated against human labels before it is trusted
Lost in the middle	—	The tendency of models to attend less reliably to material in the middle of a long context. Why more retrieved chunks can make answers worse

M

Term	Expansion	Meaning
MAP	Mean Average Precision	Average precision across all relevant results, averaged over queries. Order-sensitive
Metadata	—	Structured fields stored with a chunk: source, page, date, author, tenant, ACL. Enables filtering and citation
Metadata filtering	—	Restricting search to chunks matching structured conditions. Pre-filtering applies it during search; post-filtering after — the latter can starve results
Modular RAG	—	The generation of architectures where retrieval is conditional, repeatable, and routable rather than a fixed stage
MRR	Mean Reciprocal Rank	`1 / rank of the first relevant result`, averaged over queries. Strongly rewards ranking the answer first
Multi-hop	—	A question requiring several linked lookups, where the second search depends on the first result
Multi-tenancy	—	Serving multiple customers from one system. Requires tenant ID in every filter and every cache key

N

Term	Expansion	Meaning
Naive RAG	—	Retrieve once, generate once. The baseline — and the right starting point for every project
nDCG	Normalized Discounted Cumulative Gain	Ranking quality with graded relevance, discounted by rank and normalized to 0–1. The most informative retrieval metric
Normalization	—	Scaling a vector to unit length, which makes dot product and cosine similarity equivalent
nprobe	—	IVF parameter: how many clusters to search. Higher means better recall, slower queries

O

Term	Expansion	Meaning
OCR	Optical Character Recognition	Extracting text from images and scanned documents. Required for scanned PDFs, which contain no text layer
Overlap	—	Characters or tokens repeated between adjacent chunks, so facts near a boundary survive in at least one chunk intact

P

Term	Expansion	Meaning
Parent-document retrieval	—	Search over small precise chunks, then return the larger parent section for context. Escapes the chunk-size trade-off
PII	Personally Identifiable Information	Data identifying an individual. Detect and redact at ingestion, before embedding
Precision@k	—	What fraction of the top k results are relevant? The metric that matters after reranking
Pre-filtering / post-filtering	—	Applying metadata filters during ANN search versus after it. Post-filtering can return far fewer results than requested
Prompt injection	—	Text engineered to override a model's instructions. In RAG, the dangerous variant arrives inside retrieved documents

Q

Term	Expansion	Meaning
Quantization	—	Storing vectors at lower precision to cut memory 4–32×, at a small recall cost. Scalar: fewer bits per number. Binary: one bit per dimension. Product (PQ): compressed subvector codes
Query decomposition	—	Splitting a multi-part question into sub-questions searched separately
Query expansion	—	Generating additional phrasings of a query and searching with all of them
Query pipeline	—	The online path: understand, retrieve, rerank, assemble, generate. Runs on every question
Query rewriting	—	Turning a messy or context-dependent question into a clean standalone one. Essential for conversational RAG

R

Term	Expansion	Meaning
RAG	Retrieval-Augmented Generation	Searching your own documents for relevant passages and having a model answer using only that text
RAG-Fusion	—	Query expansion plus RRF: several rewrites, searched separately, rankings fused
RAG triad	—	Three checks: context relevance, groundedness, answer relevance
RAGAS	Retrieval-Augmented Generation Assessment	A widely used open-source RAG evaluation framework
Recall@k	—	What fraction of all relevant chunks appear in the top k? The metric that matters before reranking
Recursive chunking	—	Splitting on a priority list of separators — paragraphs, then sentences, then words — until chunks fit. The sensible default
Reflection tokens	—	Special markers a Self-RAG model emits to judge its own process: is retrieval needed, is this passage relevant, is this claim supported
Reranking	—	Re-scoring retrieved candidates with a slower, more accurate model. See Reranking & Context Assembly
Retrieval	—	The "R" in RAG: finding relevant chunks for a query
RRF	Reciprocal Rank Fusion	Merging ranked lists by summing `1 / (k + rank)` per document. The standard way to fuse hybrid search results — it needs no score calibration between the two systems

S

Term	Expansion	Meaning
Semantic cache	—	Serving a cached answer when a similar query arrives. Effective and risky — similar queries can have opposite answers
Semantic chunking	—	Splitting where meaning shifts, detected by embedding similarity between adjacent sentences
Self-RAG	—	An architecture where the model grades its own retrieval and output, and may skip retrieval entirely
Sparse retrieval	—	Keyword-based search over term-frequency vectors that are mostly zeros. BM25 is the standard. Excellent at exact terms
SPLADE	Sparse Lexical and Expansion model	A learned sparse retriever — keyword-style matching with model-learned term expansion
Speculative RAG	—	Generating several drafts in parallel with a small model, then having a large model verify and select
Stuffing	—	Placing retrieved chunks directly into the prompt. The standard augmentation method

T

Term	Expansion	Meaning
Temperature	—	Randomness in token selection. 0–0.3 for RAG — you want faithful reproduction, not creativity
Token	—	The unit models read and bill by. Roughly ¾ of a word in English
top-k	—	The k highest-scoring results from a search
Two-stage retrieval	—	Retrieve wide and fast, then rerank narrow and accurate. The standard production pattern

V

Term	Expansion	Meaning
Vector	—	An ordered list of numbers. An embedding is a vector
Vector database	—	A store that holds embeddings and finds the closest ones to a query quickly. Examples: Pinecone, Weaviate, Qdrant, Milvus, Chroma, pgvector
Vector index	—	The data structure that makes similarity search fast — HNSW, IVF, and others

The Terms Most Often Confused

Pairs worth keeping straight

Recall vs Precision

Recall = of everything relevant, how much did we find? Precision = of what we returned, how much is relevant? Optimize recall at the retrieval stage and precision after reranking. Pushing both at once with one k is how people end up satisfied with neither.

Bi-encoder vs Cross-encoder

A bi-encoder embeds query and document separately, so document vectors are precomputed and search is fast — this is your retriever. A cross-encoder reads both together and cannot precompute anything, so it is far more accurate and far too slow for a whole corpus — this is your reranker.

Faithfulness vs Correctness

Faithfulness = the answer matches the retrieved context. Correctness = the answer matches reality. An answer faithfully reproducing an outdated document is fully faithful and wrong — which is why faithfulness alone is not enough.

Dense vs Sparse retrieval

Dense matches meaning and handles paraphrase; it fails on rare exact tokens like SKUs. Sparse matches words and handles exact terms; it fails on paraphrase. They fail on opposite inputs, which is exactly why hybrid search works.

RAG vs Fine-tuning

RAG supplies facts at query time and is updated by changing documents. Fine-tuning changes weights and teaches style, format, and skill. They solve different problems and are frequently used together — fine-tune for behaviour, retrieve for knowledge.

Pre-filtering vs Post-filtering

Pre-filtering restricts the search space before ANN traversal, so you get k results that satisfy the filter. Post-filtering searches first and discards non-matching results afterwards, which can leave you with almost nothing when the filter is selective.

Where to Go Next

You have completed the RAG fundamentals. The natural next step is to build something:

Path	Start here
Build your first RAG system	Intelligent Document Q&A
Add the production techniques	Hybrid Search and RAG with Reranking
Study the advanced architectures	Agentic RAG, Graph RAG
See complete production systems	Case Studies

Glossary

How to use this page

Every term and abbreviation used anywhere in this section, expanded and explained in plain language. Organised A–Z. Where a concept has a full explanation elsewhere, the entry links to it.

Property	Value
Level	All levels — reference
Reading time	Reference; skim or search
Prerequisites	None
You will understand	Every piece of RAG vocabulary you are likely to encounter

A

Term	Expansion	Meaning
ACL	Access Control List	Metadata on a chunk recording who may see it. Must be applied as a retrieval-time filter, not after retrieval
Adaptive RAG	—	Routing each query to the cheapest sufficient path based on its complexity. See Architectures
Agentic RAG	—	Retrieval exposed as a tool an agent may call repeatedly on its own initiative. Powerful, slow, expensive
ANN	Approximate Nearest Neighbour	Search that finds almost the closest vectors, trading a little accuracy for an enormous speed gain. What makes vector databases fast
Answer Relevance	—	Evaluation metric: does the answer address the question actually asked?
Attribution marker	—	The `[1]`, `[2]` labels attached to chunks in the context so the model can cite them. Without them, citations get invented
Augmentation	—	The "A" in RAG: inserting retrieved text into the prompt

B

Term	Expansion	Meaning
Bi-encoder	—	A model that embeds query and document separately, so document vectors can be precomputed. Fast, less accurate. Contrast cross-encoder
BM25	Best Matching 25	The standard keyword-ranking algorithm. Scores documents on term frequency, adjusted for term rarity and document length. Excellent at exact and rare terms
Blue/green migration	—	Building a new index alongside the old one and cutting over, so rollback stays possible. The only safe way to change embedding models

C

Term	Expansion	Meaning
Chunk	—	One retrievable piece of a document. The unit that gets embedded, stored, searched, and inserted into the prompt
Chunking	—	Splitting documents into chunks. See Chunking Strategies
ColBERT	Contextualized Late Interaction over BERT	Late-interaction retrieval: keeps a vector per token rather than per chunk, matching at token level. More accurate than bi-encoders, much larger storage
Context Precision	—	Are the relevant chunks ranked above the irrelevant ones in the retrieved set?
Context Recall	—	Of the facts needed for the correct answer, how many are present in the retrieved context?
Context window	—	The maximum tokens a model can process in one call — prompt plus output
Contextual retrieval	—	Prepending a short generated summary of the surrounding document to each chunk before embedding, so the chunk is interpretable standalone
Corpus	—	The full collection of documents your system can retrieve from
Cosine similarity	—	Similarity between two vectors measured as the angle between them, ignoring magnitude. The standard measure for text embeddings. `1.0` identical, `0` unrelated
CRAG	Corrective RAG	Grades retrieved documents before generating, and falls back — often to web search — when the grade is poor
Cross-encoder	—	A model that reads query and document together and scores their relationship. Far more accurate than a bi-encoder, far too slow to run over a whole corpus. The standard reranker

D

Term	Expansion	Meaning
Dense retrieval	—	Search by embedding similarity — matching by meaning. Contrast sparse retrieval
Dimensions	—	The length of an embedding vector, e.g. 384, 768, 1536. More dimensions capture more nuance and cost more storage and compute
Distance metric	—	How vector closeness is computed: cosine, dot product, or Euclidean (L2)
Document ID	—	A stable identifier linking chunks back to their source document. Required to delete superseded chunks on update
Dot product	—	A similarity measure that accounts for both angle and magnitude. Equivalent to cosine when vectors are normalized

E

Term	Expansion	Meaning
efSearch	—	HNSW parameter controlling how much of the graph is explored at query time. Higher means better recall and higher latency
Embedding	—	A list of numbers representing the meaning of a piece of text. Similar meanings produce similar vectors. See Embeddings Explained
Embedding model	—	The model that converts text into vectors. Query and documents must use the same one
Extractive vs abstractive	—	Extractive answers quote source text; abstractive answers rephrase it. RAG is usually abstractive with extractive grounding

F

Term	Expansion	Meaning
FAISS	Facebook AI Similarity Search	A widely-used open-source library for fast vector similarity search. A library, not a database — no persistence or filtering of its own
Faithfulness	—	Evaluation metric: is every claim in the answer supported by the retrieved context? The primary anti-hallucination measure
Fine-tuning	—	Modifying a model's weights to change its behaviour. Teaches style and skill; RAG supplies facts. Usually complementary, not alternatives
Fixed-size chunking	—	Splitting by a fixed character or token count. Simple, and blind to meaning

G

Term	Expansion	Meaning
Golden dataset	—	The labelled evaluation set: questions, ground-truth answers, and the chunk IDs that should be retrieved
Graph RAG	—	Retrieval over a knowledge graph of entities and relationships, answering questions whose answers live between documents
Grounding	—	Constraining the model to answer only from retrieved context
Groundedness	—	See Faithfulness

H

Term	Expansion	Meaning
Hallucination	—	A fluent, confident, false statement. RAG reduces the rate and makes it detectable; it never eliminates it
Hit Rate@k	—	Did any relevant chunk appear in the top k? The coarsest retrieval metric — easily flattering
HNSW	Hierarchical Navigable Small World	The most common ANN index. A layered proximity graph navigated coarse-to-fine. Excellent recall and speed; memory-hungry
Hybrid search	—	Running dense and sparse search together and fusing the rankings. The correct default for production
HyDE	Hypothetical Document Embeddings	Have the model write a fake answer, embed that, and search with it — so answer-shaped text matches answer-shaped documents

I

Term	Expansion	Meaning
IDCG	Ideal Discounted Cumulative Gain	The DCG of a perfect ranking; the denominator that normalizes nDCG
Indexing pipeline	—	The offline pipeline: load, clean, chunk, embed, store. Runs when documents change
Indirect prompt injection	—	An attack where malicious instructions are planted inside a document your system indexes and later retrieves into the prompt
Ingestion	—	Getting source documents into the system: loading, parsing, cleaning
IVF	Inverted File Index	An ANN index that clusters vectors and searches only the nearest clusters. Lower memory than HNSW; needs a training step

K

Term	Expansion	Meaning
k	—	How many results a search returns. `top-k` is the k highest-scoring results
Knowledge cutoff	—	The date after which a model's training data ends. One of the core problems RAG solves

L

Term	Expansion	Meaning
L2 distance	Euclidean distance	Straight-line distance between two vectors. Less common than cosine for text
Late chunking	—	Embedding the whole document first, then pooling token embeddings into chunks — so each chunk vector retains document-wide context
Late interaction	—	Matching at token level at query time rather than comparing single pooled vectors. See ColBERT
LLM	Large Language Model	The model that writes the final answer. Supplies language ability; your documents supply the facts
LLM-as-judge	—	Using a model to score outputs on subjective criteria like faithfulness. Must be validated against human labels before it is trusted
Lost in the middle	—	The tendency of models to attend less reliably to material in the middle of a long context. Why more retrieved chunks can make answers worse

M

Term	Expansion	Meaning
MAP	Mean Average Precision	Average precision across all relevant results, averaged over queries. Order-sensitive
Metadata	—	Structured fields stored with a chunk: source, page, date, author, tenant, ACL. Enables filtering and citation
Metadata filtering	—	Restricting search to chunks matching structured conditions. Pre-filtering applies it during search; post-filtering after — the latter can starve results
Modular RAG	—	The generation of architectures where retrieval is conditional, repeatable, and routable rather than a fixed stage
MRR	Mean Reciprocal Rank	`1 / rank of the first relevant result`, averaged over queries. Strongly rewards ranking the answer first
Multi-hop	—	A question requiring several linked lookups, where the second search depends on the first result
Multi-tenancy	—	Serving multiple customers from one system. Requires tenant ID in every filter and every cache key

N

Term	Expansion	Meaning
Naive RAG	—	Retrieve once, generate once. The baseline — and the right starting point for every project
nDCG	Normalized Discounted Cumulative Gain	Ranking quality with graded relevance, discounted by rank and normalized to 0–1. The most informative retrieval metric
Normalization	—	Scaling a vector to unit length, which makes dot product and cosine similarity equivalent
nprobe	—	IVF parameter: how many clusters to search. Higher means better recall, slower queries

O

Term	Expansion	Meaning
OCR	Optical Character Recognition	Extracting text from images and scanned documents. Required for scanned PDFs, which contain no text layer
Overlap	—	Characters or tokens repeated between adjacent chunks, so facts near a boundary survive in at least one chunk intact

P

Term	Expansion	Meaning
Parent-document retrieval	—	Search over small precise chunks, then return the larger parent section for context. Escapes the chunk-size trade-off
PII	Personally Identifiable Information	Data identifying an individual. Detect and redact at ingestion, before embedding
Precision@k	—	What fraction of the top k results are relevant? The metric that matters after reranking
Pre-filtering / post-filtering	—	Applying metadata filters during ANN search versus after it. Post-filtering can return far fewer results than requested
Prompt injection	—	Text engineered to override a model's instructions. In RAG, the dangerous variant arrives inside retrieved documents

Q

Term	Expansion	Meaning
Quantization	—	Storing vectors at lower precision to cut memory 4–32×, at a small recall cost. Scalar: fewer bits per number. Binary: one bit per dimension. Product (PQ): compressed subvector codes
Query decomposition	—	Splitting a multi-part question into sub-questions searched separately
Query expansion	—	Generating additional phrasings of a query and searching with all of them
Query pipeline	—	The online path: understand, retrieve, rerank, assemble, generate. Runs on every question
Query rewriting	—	Turning a messy or context-dependent question into a clean standalone one. Essential for conversational RAG

R

Term	Expansion	Meaning
RAG	Retrieval-Augmented Generation	Searching your own documents for relevant passages and having a model answer using only that text
RAG-Fusion	—	Query expansion plus RRF: several rewrites, searched separately, rankings fused
RAG triad	—	Three checks: context relevance, groundedness, answer relevance
RAGAS	Retrieval-Augmented Generation Assessment	A widely used open-source RAG evaluation framework
Recall@k	—	What fraction of all relevant chunks appear in the top k? The metric that matters before reranking
Recursive chunking	—	Splitting on a priority list of separators — paragraphs, then sentences, then words — until chunks fit. The sensible default
Reflection tokens	—	Special markers a Self-RAG model emits to judge its own process: is retrieval needed, is this passage relevant, is this claim supported
Reranking	—	Re-scoring retrieved candidates with a slower, more accurate model. See Reranking & Context Assembly
Retrieval	—	The "R" in RAG: finding relevant chunks for a query
RRF	Reciprocal Rank Fusion	Merging ranked lists by summing `1 / (k + rank)` per document. The standard way to fuse hybrid search results — it needs no score calibration between the two systems

S

Term	Expansion	Meaning
Semantic cache	—	Serving a cached answer when a similar query arrives. Effective and risky — similar queries can have opposite answers
Semantic chunking	—	Splitting where meaning shifts, detected by embedding similarity between adjacent sentences
Self-RAG	—	An architecture where the model grades its own retrieval and output, and may skip retrieval entirely
Sparse retrieval	—	Keyword-based search over term-frequency vectors that are mostly zeros. BM25 is the standard. Excellent at exact terms
SPLADE	Sparse Lexical and Expansion model	A learned sparse retriever — keyword-style matching with model-learned term expansion
Speculative RAG	—	Generating several drafts in parallel with a small model, then having a large model verify and select
Stuffing	—	Placing retrieved chunks directly into the prompt. The standard augmentation method

T

Term	Expansion	Meaning
Temperature	—	Randomness in token selection. 0–0.3 for RAG — you want faithful reproduction, not creativity
Token	—	The unit models read and bill by. Roughly ¾ of a word in English
top-k	—	The k highest-scoring results from a search
Two-stage retrieval	—	Retrieve wide and fast, then rerank narrow and accurate. The standard production pattern

V

Term	Expansion	Meaning
Vector	—	An ordered list of numbers. An embedding is a vector
Vector database	—	A store that holds embeddings and finds the closest ones to a query quickly. Examples: Pinecone, Weaviate, Qdrant, Milvus, Chroma, pgvector
Vector index	—	The data structure that makes similarity search fast — HNSW, IVF, and others

The Terms Most Often Confused

Pairs worth keeping straight

Recall vs Precision

Bi-encoder vs Cross-encoder

Faithfulness vs Correctness

Dense vs Sparse retrieval

RAG vs Fine-tuning

Pre-filtering vs Post-filtering

Where to Go Next

You have completed the RAG fundamentals. The natural next step is to build something:

Path	Start here
Build your first RAG system	Intelligent Document Q&A
Add the production techniques	Hybrid Search and RAG with Reranking
Study the advanced architectures	Agentic RAG, Graph RAG
See complete production systems	Case Studies

Glossary

On this page

Glossary

On this page