RAG Reference Pipeline

Fact-check status

Reviewed claim by claim on 2026-09-02 against the primary or official sources identified inline as S1 and S2. Recommendations are design guidance unless a cited standard makes them normative.

Key jargon

Term	Plain-language meaning
Retrieval-augmented generation (RAG)	Generating an answer with evidence retrieved from an external knowledge collection. S1
Retriever	The component that selects candidate passages for a query. S1
Reranker	A second-stage model or rule that reorders retrieved candidates by relevance. S1
Grounding	Constraining an answer to supplied evidence and making support inspectable. S1

Key concepts

RAG is a pipeline whose retrieval and generation stages can fail independently. S1
Evaluate source ingestion, retrieval, evidence support, and final answers separately. S1

Concept map

flowchart LR
    A["Ingest and index sources"] --> B["Retrieve candidates"]
    B --> C["Rerank and assemble evidence"]
    C --> D["Generate grounded answer"]

Diagram semantics verified against S1 and S2 on 2026-09-02.

Retrieval-augmented generation supplies selected external evidence to a generator. It can improve freshness and grounding, but it does not guarantee either. S1

flowchart LR
    D[Sources] --> I[Parse and normalize]
    I --> C[Chunk and enrich metadata]
    C --> X[Index]
    Q[Question] --> R[Retrieve]
    X --> R
    R --> K[Rerank and filter]
    K --> P[Construct grounded prompt]
    P --> M[Model]
    M --> V[Validate answer and citations]

Diagram semantics verified against S1 and S2 on 2026-09-02.

Two evaluation planes

Retrieval: did the system find the evidence needed? S1
Generation: did the answer faithfully use that evidence? S1

Exercise

Create five short documents and ten questions with known supporting passages. Record retrieved passages separately from generated answers so failure location is visible.

Primary source

Lewis et al., Retrieval-Augmented Generation for Knowledge-Intensive NLP Tasks