11 · Enterprise RAG Pipeline (v2)

An alternate layout of the enterprise RAG pipeline — the same question-to-grounded-answer flow, drawn to emphasise the offline ingest and online retrieval stages side by side.

How it works

1. 1Ingest & index (offline): chunk documents, embed them and store with metadata.
2. 2Rewrite the query and run hybrid search (vector + BM25).
3. 3Filter by metadata, retrieve top-K, then re-rank with a cross-encoder.
4. 4Expand context with parent/neighbouring chunks.
5. 5Generate the grounded answer with the LLM and return citations.

Key points

• Same pipeline as the primary diagram, different visual emphasis.
• Separates the offline indexing stage from the online answering stage.
• Reciprocal Rank Fusion (RRF) merges keyword and vector results.
• Access Control Lists (ACL) enforce who can retrieve what.

Why it matters

Seeing ingest and retrieval side by side clarifies that RAG has two halves: preparing the knowledge base offline and answering questions online.

Frequently asked questions

What is RRF?

Reciprocal Rank Fusion — a way to merge keyword and vector search rankings into one combined result.

Why separate offline and online?

Indexing happens once (offline); answering happens per query (online). They have different performance needs.