RAG pipeline icons

The 40 icons, described

Ingestioningestion

Ingestion — getting the source material loaded into a pipeline or a knowledge base.

Ingestion pipelineingestion-pipeline

An ingestion pipeline — documents go in, index entries come out.

Document parsedocument-parse

Document parsing — reading the structure out of a file and converting it to usable data.

Document splitdocument-split

Document split — one document cut into pieces, the first step of chunking for retrieval.

Chunkchunk

A chunk — one piece of a document, split to a size that suits retrieval and embedding.

Chunkingchunking

Chunking — the whole document made into pieces that fit the context.

Chunk overlapchunk-overlap

Chunk overlap — the stretch of text two neighbouring chunks both contain so context is not lost.

Chunk sizechunk-size

Chunk size — how big each piece of a split document is, measured in tokens or characters.

Embeddingembedding

A vector embedding — a piece of meaning turned into coordinates a model can compare and search.

Embed batchembed-batch

Embed batch — many texts through the embedding model at once for throughput.

Vector databasevector-database

A vector database — stores embeddings and finds the nearest neighbours by similarity.

Vector indexvector-index

A vector index — embeddings filed for finding, an approximate nearest-neighbour store.

Vector searchvector-search

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

Index rebuildindex-rebuild

Index rebuild — the index made again from scratch in a batch reindex.

HNSWhnsw

HNSW — a layered graph index you can search for nearest neighbours quickly.

Retrieverretriever

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

Hybrid searchhybrid-search

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

Keyword searchkeyword-search

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

Semantic searchsemantic-search

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

Rerankerreranker

A reranker — a model that puts retrieved results back in the order of true relevance.

Cross-encodercross-encoder

A cross-encoder — a reranker that judges query and document together rather than apart.

Top-ktop-k

Top-k — keep the k best candidates and drop the rest before sampling.

Relevance scorerelevance-score

A relevance score — how close a result is to what was actually asked.

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.

Groundinggrounding

Grounding — an answer tied to a real source so every claim can be traced to evidence.

Citationcitation

A citation — where that claim came from, a reference to the source.

Evidenceevidence

Evidence — the passage that supports a claim, the proof behind an answer.

Answer synthesisanswer-synthesis

Answer synthesis — several retrieved sources composed into one written answer.

RAG evaleval-rag

RAG eval — grading the retrieval as well as the answer it produced.

Groundednessgroundedness

Groundedness — every claim standing on a source, verified and solid.

Freshnessfreshness

Freshness — how recently the data was checked or updated, its age before it goes stale.

Corpuscorpus

A corpus — the whole body of documents a system has read and can draw on.

Knowledge baseknowledge-base

A knowledge base — the body of documents and facts a system can draw on for answers.

Knowledge graphknowledge-graph

A knowledge graph — facts joined to other facts as entities and edges.

Multi-hopmulti-hop

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

Query expandquery-expand

Query expansion — one question broadened into several searches.

HyDEhyde

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

Re-embedreembed

Re-embed — a new model over the same corpus, every vector computed again.

Deduplicatededupe

Deduplicate — two of the same, with one of them dropped so only unique items remain.

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