Retrieval icons

The 75 retrieval icons in RAG & Search, described

Agentic RAGagentic-rag

Agentic RAG — the agent decides what to look up next in a retrieval loop.

Answer relevanceanswer-relevance

The lens with a speech bubble in it — whether the answer speaks to the question.

Bi-encoderbi-encoder

The lattice with two towers side by side — query and document embedded apart, compared after.

Chunk metadatachunk-metadata

A page with a line and a tag under it — what is known about a chunk besides its text.

Citation spancitation-span

A page with a line and a stretch of the next one marked off — exactly which words an answer rests on.

ColBERTcolbert

The lattice with three bars side by side — one vector per token, kept apart.

Vector collectioncollection-vector

The lattice with a box in it — the set of vectors an index holds.

Context assemblycontext-assembly

Context assembly — the retrieved pieces gathered and composed into one prompt.

Context budgetcontext-budget

A context budget — how much room is left in the window for more tokens.

Context injectioncontext-injection

Context injection — putting extra material into the prompt before the model sees it.

Context overflowcontext-overflow

Context overflow — more tokens than the window will hold, so something gets truncated.

Context relevancecontext-relevance

The lens with brackets in it — whether what was fetched was worth fetching.

Contextual chunkcontextual-chunk

A page with a header over a box — a chunk sent with the context that explains it.

Delete vectordelete-vector

The lattice with a cross in it — a vector taken out of the index.

Dense searchdense-search

Dense search — retrieval by embedding vector rather than by matching words.

Dense vectordense-vector

The lattice with two full lines — a vector with something in every dimension.

Faithfulnessfaithfulness

The lens with a shield in it — whether the answer says only what the sources say.

Metadata filterfilter-metadata

A metadata filter — narrow the candidates by tags or facets before you search.

Filter resultfilter-result

A filter result — what made it through the filter, the matching output.

Fusion rankfusion-rank

The funnel with two columns tied by a bar — two result lists folded into one.

HNSW graphhnsw-graph

The lattice with a tree in it — layers of links, coarse on top and fine below.

Hybrid fusehybrid-fuse

Hybrid fuse — keyword and semantic results merged into one ranked list.

Hybrid searchhybrid-search

Hybrid search — keywords and meaning combined, lexical and vector results fused.

Hybrid weighthybrid-weight

The funnel with a big and a small weight under it — how much keyword and how much vector.

HyDEhyde

HyDE — imagine a hypothetical answer and search with its embedding.

Incremental indexincremental-index

The lattice with a plus in it — new vectors added without rebuilding the rest.

Index refreshindex-refresh

The lattice with an arrow looping back — the index rebuilt from the source.

Index snapshotindex-snapshot

The lattice with an eye in it — the index as it was at one moment, kept.

IVF listivf-list

The lattice with a centroid between two walls — vectors grouped into cells, each with its list.

Keyword searchkeyword-search

Keyword search — find the literal words themselves, exact text matching.

Late interactionlate-interaction

The lattice with two points tied together — query and document matched token by token, at the end.

Metadata filtermetadata-filter

The funnel with a tag under it — only chunks whose fields match.

MRRmrr

The lens with a point ahead of a line — how high up the first right answer sits.

Multi-hopmulti-hop

Multi-hop — the answer is two documents away, reached by chaining retrievals.

Multi-querymulti-query

The lens with three lines in it — the same question asked several ways.

Vector namespacenamespace-vector

The lattice with brackets in it — vectors kept apart by whose they are.

NDCGndcg

The lens with bars stepping down — results graded by how good they are and how high they sit.

Parent documentparent-document

A page with a small box in it — the whole document a chunk came from.

Passage mergepassage-merge

A page with two lines joining into one — neighbouring chunks stitched back together.

Post-filterpost-filter

The funnel with a list under it — the results, thinned after they were found.

Post-filterpostfilter

Post-filter — search wide, then throw most of the results back.

Precision@kprecision-at-k

The lens with a ring in it — how much of the top results is worth having.

Pre-filterprefilter

Pre-filter — narrow the candidates first, then search what is left.

Product quantisationproduct-quantisation

The lattice cut into strips — a vector split into parts, each stored as a code.

Quantise vectorquantise-vector

The lattice with a staircase in it — a vector stored in fewer bits.

Query decompositionquery-decompose

The lens with one line splitting into two — a question broken into the questions inside it.

Query embeddingquery-embed

The lattice with an arrow in it — a question turned into a vector.

Query expandquery-expand

Query expansion — one question broadened into several searches.

Query rewritequery-rewrite

The lens with a pencil stroke in it — the question reworded before it is searched.

RAG pipelinerag-pipeline

A RAG pipeline — retrieve the relevant passages, then generate with what you found.

Recall@krecall-at-k

The lens with a check in it — how much of what should be found is in the top results.

Reciprocal rank fusionreciprocal-fusion

The funnel with an arrow pointing both ways — each list's ranks traded into one score.

Recursive chunkingrecursive-chunk

A page with lines that step in — a document split, and the pieces split again.

Rerank windowrerank-window

The funnel with brackets under it — the top results, and only those, scored again.

Retrieval cacheretrieval-cache

The lattice with a bolt in it — a query answered from what was fetched last time.

Retrieval depthretrieval-depth

The lattice with an arrow going down — how far down the ranking the search reads.

Retrieval evaluationretrieval-eval

The lens with a check and a cross in it — the hits and the misses, counted.

Retrieval fallbackretrieval-fallback

The funnel with an arrow pointing back — when nothing good is found, the plainer search runs.

Retrieverretriever

A retriever — the component that fetches the documents most likely to help answer a query.

Semanticsemantic

Semantic — by what it means rather than what it says, meaning beyond the words.

Semantic searchsemantic-search

Semantic search — find things by what they mean, using vectors and similarity rather than keywords.

Sentence windowsentence-window

A page with posts through the middle line — a sentence kept with the ones around it.

Cosine similaritysimilarity-cosine

The lattice with two rays from one point — how close two vectors point.

Dot productsimilarity-dot

The lattice with a vector and its shadow on the axis — how far one vector reaches along another.

Euclidean distancesimilarity-euclidean

The lattice with two points and the straight line between — how far apart two vectors sit.

Similarity searchsimilarity-search

Similarity search — find what is most like this one, the nearest neighbours.

Source weightsource-weight

A page with a weight on it — how much one source counts against another.

Sparse searchsparse-search

Sparse search — match the words themselves, lexical and exact.

Sparse vectorsparse-vector

The lattice with one point over an empty line — a vector that is mostly zeros.

SPLADEsplade

The lattice with a wisp in it — a sparse vector the model learned to expand.

Step backstep-back

Step-back prompting — ask the bigger, more abstract question first.

Step-back promptstep-back-prompt

The lens with a chevron rising over a line — the broader question asked first.

Sub-questionsub-question

The lens with a question mark in it — one of the smaller questions a big one breaks into.

Subquestionsubquestion

A subquestion — the big ask cut into pieces that can each be answered.

Vector searchvector-search

Vector search — find the nearest points in embedding space, a k-nearest-neighbour lookup.

In code, each is one import — import { AgenticRag } from "@iconmind/react/icons/agentic-rag" — and the same name in Vue, Svelte, Solid, Preact, React Native, Astro, Blade and Flutter.

Narrow it by tag