Skip to content

Back to the catalog

Mapping the RAG Landscape: A Four Axis Taxonomy of Efficiency, Defense, Interactivity, and Reasoning
arxiv.orgpaper

Mapping the RAG Landscape: A Four Axis Taxonomy of Efficiency, Defense, Interactivity, and Reasoning

This arXiv survey organizes recent Retrieval-Augmented Generation (RAG) research into a four-axis taxonomy covering retrieval efficiency, robustness/security, user-driven interactivity, and multi-step reasoning. It formalizes RAG framework components, reviews dense/sparse retrieval, fusion strategies, embedding optimization, and RL-based retrieval policies, and compares Naive, Advanced, and Modular RAG architectural variants alongside evaluation practices and domain applications. The paper closes by outlining open challenges in retrieval quality, reliability, domain adaptation, scalability, and explainability.

TL;DR A four-axis taxonomy survey mapping the RAG landscape across efficiency, defense, interactivity, and reasoning.
Open original ↗
Document

What it claims

The survey sorts RAG work into four axes. Each axis has a seed paper and a research question:

AxisSeed paperQuestion it asks
Compression and efficiencyRAPTORWhich retrieval and compression strategies best trade factual accuracy against inference cost on large, mixed corpora?
Defensive RAGSELF-RAGWhat vulnerabilities does retrieval add, and how well do current defences mitigate them?
Interactive, user-centric RAGChat-RECHow can personalization from user intent, history and preferences be balanced against factual precision?
Complex reasoningGenGroundDoes iterative, step-conditioned retrieval beat single-pass retrieval on multi-hop tasks, and what stops errors from piling up across steps?

The vulnerabilities named under the defensive axis are corpus poisoning, demographic bias amplification and privacy leakage.

The authors say these axes are "orthogonal and under-reviewed." Earlier surveys, they argue, focused on dense-vs-sparse comparisons or generic pipelines. The paper maps the questions forward through the document:

  • §3 covers what has been achieved on each question.
  • §5 covers what remains unresolved.
  • §6 covers promising directions.

How it works

The literature review follows Snyder's semi-systematic framework and is "PRISMA 2020-inspired."

  • Sources: Google Scholar, Semantic Scholar and arXiv.
  • Coverage window: January 2020 to January 2025.
  • Search terms: keywords such as "RAG defense", "RAG fairness" and "multi-hop retrieval LLM".
  • Initial pool: about 780 records.
  • Included: work that integrates retrieval with generation, proposes RAG benchmarks, or studies security, fairness or safety.
  • Excluded: standalone retrievers, standalone LLMs, and work with no technical or empirical contribution.

A citation-snowballing pass started from Lewis et al. 2020 plus the four axis seeds. It traced both backward and forward citations and added:

  • HippoRAG (efficiency)
  • RuleRAG (defence)
  • MemoCRS (interactive)
  • follow-ups to GenGround (reasoning)

The formal section covers four pieces:

  • BM25 scoring.
  • DPR as an inner product of query and document encoders.
  • Softmax-normalized retrieval scores.
  • Marginalization over the retrieved documents: P(y|x) = Σ P(y|x,d)·P(d|x).

The paper explicitly says this softmax marginalization is a modelling choice, not something every RAG system requires. Because P(d|x) is normalized only over the top-k retrieved set, it is not a calibrated relevance estimate over the whole corpus. That caveat applies to anyone who reads retrieval scores as confidence. The fetched text cuts off mid-argument here, so any further costs the authors list are not visible.

Limits

  • Coverage ends January 2025. The paper was posted on 1 Oct 2026, so roughly 20 months of RAG work fall outside the systematic search. The snowballing pass was "scoped as a targeted completeness check" and did not re-run the identification stage.
  • Final paper count unknown. The fetched text gives the ~780 identified records but not how many were included after screening.
  • No original experiments. The visible text contains no new benchmarks or comparative numbers.
  • Partial view of the paper. The page text ends inside §3.2. Sections 4–6 (evaluation practices, applications, open challenges and directions) were not available, so nothing here describes their content.

Technologies named here

  • Retrieval-Augmented Generation

Techniques named here

  • RAG
  • Dense retrieval
  • Sparse retrieval
  • Fusion strategies
  • Embedding optimizations
  • Reinforcement learning based retrieval policies
  • Naive RAG
  • Advanced RAG
  • Modular RAG

Related in llm-apps