How to Read RAG Papers: From Dense Retrieval (DPR) to Lewis RAG

Bookmark this page. The paper-reading hub already has three reading paths. If you have just finished DPR and Lewis RAG, this page shows how those 2020 classics connect and which note to open next.

This is not a new paper note, and it does not replace the six Paper Essence questions in each linked article. It only answers how the nodes connect, which control point changed, and which link to open next. The sibling orientation page is the How to Read AI Agent Papers: From CoT and WebGPT to ReAct (Agent vs Retrieval).

Ninety-second mental model

1. Sparse first-stage retrieval (BM25 / TF-IDF) is the default DPR has to beat. The later BM25 at scale note studies accuracy–cost curves as corpora grow, not the 2020 open-domain QA protocol.
2. DPR (Karpukhin et al., 2020) changes the first stage to a dense dual encoder (NQ top-20 78.4 vs BM25 59.1; extractive Exact Match 41.5). Lewis RAG (Lewis et al., 2020) changes generation to condition on retrieved Wikipedia passages $z$ (NQ RAG-Seq Exact Match 44.5). Do not mix those two tables.
3. The 2025–26 notes extend the sequence into multimodal retrieval, tool routing, graphs, memory adaptation, evidence discovery, financial hard negatives, set-wise reranking, and read-before-final. They use different tasks and evaluations, so their numbers do not belong in 2020 tables.
4. 2020 RAG is not a production RAG platform and not an agent loop. It has no enterprise ACL, hybrid stack, citation-service specification, or search / read / final agent cycle.

REALM’s expensive joint-pretraining control point now has a deep read: REALM. ORQA remains arXiv-only (arXiv:1911.03868); do not invent a fake note. Self-RAG’s when-to-retrieve control point has a deep read: Self-RAG.

The method map

In the diagram, “on this site” means that a full note is available here. REALM now has an on-site deep read, ORQA remains arXiv-only related work, and Self-RAG covers when to retrieve.

flowchart TB
  Sparse["sparse retrieval BM25 / TF-IDF"]
  REALM["on this site: REALM 2020<br/>expensive joint pre-training"]
  Sparse --> DPR["DPR 2020<br/>dense dual-encoder passage retrieval"]
  REALM --> DPR
  DPR --> RAG["Lewis RAG 2020<br/>retrieval conditions generation"]
  Sparse --> BM25leaf["on this site: BM25 at scale"]
  RAG --> Anything["on this site: RAG-Anything"]
  RAG --> RAGMCP["on this site: RAG-MCP"]
  RAG --> Graph["on this site: GraphRAG vs RAG"]
  RAG --> ERM["on this site: RAG without Forgetting"]
  RAG --> DocMemo["on this site: DocMemo"]
  RAG --> FinRank["on this site: FinRank"]
  RAG --> Rubric["on this site: RubricRanker"]
  RAG --> ReadGate["on this site: before reasoning fails"]
  RAG --> SelfRAG["on this site: Self-RAG<br/>when to retrieve"]

How to walk the map

Path A · Fastest: DPR, then Lewis RAG

1. DPR: lock the dense dual-encoder contract that replaces sparse first-stage retrieval.
2. Lewis RAG: see how retrieved $z$ conditions generation (RAG-Sequence / RAG-Token).
3. Stop. Read later work only when you need the question it addresses.

Path B · Full classic sequence: sparse opponent → REALM → DPR → RAG; ORQA stays arXiv-only

Follow REALM → the sparse BM25 comparison inside the DPR note → DPR → Lewis RAG. ORQA remains related prior work only: ORQA arXiv:1911.03868. This sequence establishes the method foundations; the hub’s retrieval-systems path then adds multimodal, tool, graph, and runtime work.

Path C · Pick the next paper from the job

Where the work is stuck	Start with this paper	Core question it extends
Growing corpora, hybrid / scale cost	BM25 at scale	Sparse accuracy–cost at scale, not the 2020 ODQA protocol
PDFs, figures, multimodal documents	RAG-Anything	Retrieval targets become referable multimodal nodes, not text-only passages
Too many tool schemas; retrieve then call	RAG-MCP	Retrieval used for tool-description routing, not authorization
Whether a knowledge graph is worth it	GraphRAG vs RAG	Whether graph indexing is a necessary upgrade for this problem class
Whether successful expansions can write back	RAG without Forgetting	Memory write-back behind a correctness gate
Evidence needs dynamic discovery, not one top-$k$	DocMemo	The evidence-discovery process
Finance / hard negatives, evidence grounding	FinRank	Ranking and hard negatives
Set-wise rerank for deep research	RubricRanker	The reranking contract
The system searched but answered before reading evidence	Before Reasoning Can Fail	Read-before-final; Read-Gate is not a substitute for retrieval quality
Whether the model should decide when to retrieve	Self-RAG	when-to-retrieve; reflection tokens, not Read-Gate

Node table: control point, one sentence, do-not-misread

Node	Control point changed	One sentence	Link	Do not misread
Sparse BM25 / TF-IDF	Whether the first stage is lexical inverted index	The default sparse retrieval DPR must beat	See contrast inside the DPR note	Opponent baseline, not the on-site BM25-at-scale leaf
REALM	Whether pre-training jointly retrieves and asynchronously refreshes an index	Costly retrieval-augmented pre-training; DPR argues you need not pay that bill	Already on this site	Early method; not production RAG and not the Lewis RAG generation table
ORQA	Latent dense retrieval + ICT (related prior)	One direct comparison starting point for REALM; no on-site deep read	arXiv:1911.03868	arXiv-only; do not expect a 2026-format note
DPR	Whether the first stage becomes a dense dual encoder	Two BERTs, dot-product MIPS; NQ top-20 78.4 vs 59.1; extractive EM 41.5	Already on this site	The retrieval table is not Lewis RAG’s generation table
Lewis RAG	Whether generation conditions on retrieved $z$	BART plus a DPR-initialized retriever; NQ RAG-Seq 44.5	Already on this site	2020 method paper ≠ production RAG platform ≠ agent loop
BM25 at scale	How accuracy–cost bends as corpora grow	Large-corpus sparse retrieval and agent cost	Already on this site	Later scale study; not the 2020 ODQA protocol
RAG-Anything	Whether document nodes keep multimodal structure	Tables / figures / formulas stay referable; not caption-only	Already on this site	Later multimodal method; its numbers do not belong in 2020
RAG-MCP	How to choose among too many tool schemas	Retrieve candidate schemas, then let the executor call	Already on this site	Later routing method; retrieval is not authorization
GraphRAG vs RAG	Whether multi-hop / structural work needs a graph	Systematic comparison of graph vs vector adoption boundaries	Already on this site	Later comparison; not a drop-in replacement for Lewis RAG
RAG without Forgetting	Where successful expansions are written back	Index adaptation behind a correctness gate	Already on this site	Later adaptation method; write-back is not unconditional memory
DocMemo	Whether evidence is fetched once or discovered dynamically	Dynamic evidence discovery, not a fixed top-$k$	Already on this site	Later evidence method; not xMemory, which is agent memory
FinRank	How evidence is ranked under financial hard negatives	Evidence-grounded ranking	Already on this site	Later ranking method; its numbers do not belong in 2020
RubricRanker	How deep research does set-wise rerank	Set-wise reranking with rubrics	Already on this site	Later reranker; not first-stage retrieval
Before reasoning can fail	Whether a read happens after search and before final	A pre-evidence procedural failure, not a wrong answer after reading gold	Already on this site	Later procedural study; Read-Gate is not a substitute for retrieval quality
Self-RAG	Whether / when to call retrieval	Self-reflection decides retrieve / critique	Already on this site	when-to-retrieve; not Read-Gate and not a production RAG gate

What this page is not

• It does not replace the six Paper Essence questions. Each linked note still has to carry: the problem, why the prior approach was insufficient, the core idea, how one input moves through the method, which evidence supports the headline, and where the claim stops. This page only orients.
• It does not invent an ORQA note. ORQA stays arXiv-only. REALM and Self-RAG now have on-site deep reads (REALM, Self-RAG); this page only orients and does not rewrite the six Paper Essence questions.
• It does not rewrite the hub’s retrieval-systems path narrative. That path starts at REALM → DPR → RAG → Self-RAG and then adds multimodal, tool, graph, and runtime work. This page remains an orientation guide, not a fourth path type.
• It does not back-fill later numbers into classics, and it does not mix the DPR and Lewis RAG tables. Evidence, author claims, and Bloss0m judgment stay in the individual notes.
• It does not force xMemory or AskChem into this sequence. xMemory is agent memory, while AskChem is claim-centered synthesis; neither belongs on this RAG method path.

If the reading method itself is still unfamiliar, pair this map with Efficient Academic Paper Reading: The Three-Pass Approach. If you want product architecture rather than a paper family, start from the Enterprise RAG guide. For the Agent-family sibling map, see the How to Read AI Agent Papers: From CoT and WebGPT to ReAct.

How to use this guide

• Entering from the paper-reading hub: the three PATHS remain; if you need the DPR / Lewis RAG method map, stop here and follow a link. Retrieval-systems path: #reading-paths.
• Entering from a classic note: if the article links to a “reading map,” it means this page.
• Traditional Chinese edition: use the language toggle on this page.

References

• Paper-reading hub (three PATHS; retrieval-systems path at #reading-paths)
• Karpukhin et al., 2020, Dense Passage Retrieval
• Lewis et al., 2020, Retrieval-Augmented Generation
• Guu et al., 2020, REALM — on-site note
• Guu et al., 2020, REALM arXiv
• Lee et al., 2019, ORQA
• Asai et al., 2023, Self-RAG
• Method post on this site: Three-pass reading
• Sibling orientation page (Agent family): How to Read AI Agent Papers: From CoT and WebGPT to ReAct
• A different map (Harness blogs, not this paper family): How to Read Harness Engineering: Setup and Verification for Long-Running Agents