Information Retrieval
Information retrieval (IR) is the process of obtaining relevant data or content from large datasets or information repositories in response to a user's query[^c1]. It is a cornerstone of technologies such as search engines, digital libraries, databases, and recommendation systems[^c2], and it underpins how people locate the specific content they need — whether text, images, audio, or video — within collections that grow continuously larger.
The field is organized around a small set of foundational ideas. Core concepts such as the information need, the query, and relevance define what retrieval systems are trying to achieve. Indexing structures such as the inverted index make searching efficient, while term-weighting schemes and retrieval models — the Boolean model, the vector space model, probabilistic and language-modeling approaches, and neural retrieval — determine how documents are matched against queries and ranked. Evaluation methodology, built on measures such as precision, recall, mean average precision, and NDCG, and on the test collections produced by evaluation campaigns such as TREC, provides the infrastructure for measuring system effectiveness[^c4].
IR is deployed across a wide range of applications, from web search and question answering to recommender systems, cross-lingual retrieval, and legal and [[Enterprise Search|enterprise search]]. The field's current frontier is shaped by artificial intelligence and deep learning[^c3]: neural models have made dense and learned retrieval practical, and large language models have introduced retrieval-augmented generation, renewing the connection between classical search and modern generative systems while raising new questions about evaluation and the future of information access. A standard pattern in this landscape is hybrid retrieval, which pairs keyword-based full-text search with vector search[^c13], and agentic search has consolidated its position as a distinct paradigm for complex information seeking[^c9]. Retrieval-augmented generation has also extended beyond text: multimodal RAG incorporates multiple modalities such as text, images, audio, and video to enhance the generated outputs[^c15]. The shift is visible in mainstream web search, where AI-generated answers and conversational modes have been merged into a single integrated search experience[^c5], and in agentic systems that interleave retrieval, reasoning, and ranking. These systems expose a persistent gap between benchmark performance and user experience, as agents score well on standard search benchmarks yet real users consistently find results unsatisfying[^c8]. At the same time, the value of classical retrieval is being reasserted: studies have shown that simple pattern matching can suffice for agentic search[^c18], renewing attention to Boolean and lexical methods inside LLM-based systems.
The evaluation infrastructure of the field is being renegotiated alongside these changes. Large language models are increasingly used as automated relevance judges, prompting the development of meta-evaluation resources that compare LLM judges against human assessments and probe vulnerabilities such as circularity, overfitting, self-preference, and content manipulation[^c7]. The judges themselves are also under scrutiny: LLM judges vary substantially in performance across tasks and evaluation aspects[^c16], and retrieval quality must be measured in isolation from the generative model it feeds into[^c17]. The reliability of the benchmarks themselves is also being questioned: standard benchmarks disagree substantially about model rankings, with average cross-benchmark agreement of only 0.52 in Kendall's Tau[^c10], a discrepancy that a train-before-test evaluation protocol substantially narrows[^c11]. As benchmarks saturate, a further line of work argues that saturation is not only a property of the task set but also of the evaluation protocol[^c20], prompting adaptive judging schemes that recover ranking signal from near-tied scores. The major evaluation campaigns continue to evolve as well: TREC's 2025 call for participation announced that TRECVID and TAC had been folded back into TREC[^c6], with TRECVID's multimedia tracks running under the TREC umbrella since the 2024 cycle[^c14], though TAC's own archive lists its last workshop entry as 2023 with no 2024 cycle[^c12]. In 2026 TREC's RAG track introduced a new corpus, ClimbMix-400b, replacing MS MARCO v2.1[^c19], and billed itself as the first agent-first track at the conference, extending the test-collection model to systems that retrieve and generate.