Forgotten History or Test-of-Time? Retrospect and Prospect on RAG from an IR Perspective
Xiaoyan Zhao, Yujie Cai, Yang Zhang, Grace Hui Yang, Tat-Seng Chua
cs.AI
2026-08-09
This survey traces RAG's core ideas to decades of pre-LLM IR and QA research, arguing the LLM is a new interface layer atop an old QA architecture, not the origin point.
RAG is widely treated as a new paradigm born to fix large language models, a mechanism that grounds their output in external knowledge. This survey argues that narrative is incomplete. RAG's core ideas, including combining retrieval with generation, knowledge augmentation, answer verification, and iterative query refinement, were studied and instantiated in IR (information retrieval) and QA (question answering) research well before LLMs, dating to the early 2000s. The authors set out to map this overlooked lineage systematically and explain why it has gone unrecognized.
The argument is historical tracing plus systematic comparison. The authors chase modern RAG and Agentic RAG back along four threads:
The authors single out QUALIFIER (2002-2003) as proto-Agentic RAG: it ran answer-verification-driven iterative query refinement with a Successive Constraint Relaxation strategy and ranked 2nd at TREC 2002, answering 290 of 500 questions (58%), behind LCC (83%) but ahead of IBM, MIT, USC/ISI, and BBN. Its design principle, tighten constraints first to maximize precision then relax iteratively if no answer is found, is exactly what today's Agentic RAG does.
This is a position/survey paper with no experiments; the conclusions are argumentative rather than numeric. It offers two usable outputs:
First, a conceptual-continuity framework (paper Figure 1): classical QA leads to modern RAG (single-pass retrieve-then-answer), while QUALIFIER's iterative retrieve-refine line leads to Agentic RAG, each with its own antecedent.
Second, four directions covered by prior work but underused by today's RAG (paper Figure 2): personalized RAG (user long-term-preference modeling), proactive RAG (helping clarify latent needs, not just answering), governance-aware RAG (structured guardrails, access control, auditability), and user-centered evaluation (metrics that resist feedback bias).
The authors also explain why this lineage went unrecognized: fragmentation between the IR and RAG communities, terminology drift across eras, and recency bias in a fast-moving field.
For people building RAG systems, the value here is avoiding redundant work. The paper surfaces a set of mechanisms (user modeling, answer validation, query refinement) that older QA systems already validated and that today's RAG is at risk of reinventing, and points back to the IR literature to borrow from. Positioning the LLM as a new interface over an old architecture is also a useful lens: it clarifies what is genuinely new (generation quality) versus what is an old problem reskinned (retrieval, verification, rewriting).
As a survey/position piece, every conclusion is argumentative; no experiment validates whether the four directions actually pay off, so they read more as a research agenda than proven conclusions. The authors concede the difficulty of each direction: personalization must handle sparse and noisy user signals and LLM long-context degradation; proactive RAG must balance intervention timing against interruption cost; governance-aware RAG is harder under generative uncertainty than in traditional IR; and user-centered evaluation must solve attribution across multi-stage pipelines and inconsistency in generative output. Readers must also judge for themselves how much of twenty-year-old QA experience transfers into today's neural retrieval stack, given the gap in representation and scale.