Abstract:

[Purpose/Significance] With the rapid development of large language models and web-connected search functions, AI retrieval has become an important new form of information access and is reshaping the technical logic, interaction mode, and service structure of traditional information retrieval. Taking DeepSeek as a representative case, this study compares large model-driven AI retrieval with traditional information retrieval from the perspective of information resource management. The purpose of this paper is not merely to discuss whether AI retrieval will replace traditional retrieval, but to clarify the functional boundaries, structural differences, complementary roles, and possible integration paths between the two. Compared with previous discussions that mainly focus on technical performance or application experience, this study emphasizes the transformation of information retrieval from document location to semantic understanding, evidence organization, reasoning support, and knowledge generation. It therefore provides a useful reference for understanding the evolution of retrieval services in the intelligent information environment. [Method/Process] Based on a review of the historical development of information retrieval, this paper conducts a systematic comparative analysis of traditional information retrieval and AI retrieval from several dimensions, including technical principles, information demand expression, retrieval mechanisms, result organization, application scenarios, advantages, limitations, and governance risks. Traditional information retrieval is examined through its core processes of keyword matching, controlled vocabulary, Boolean logic, index construction, relevance ranking, and document return. These mechanisms reflect a relatively mature retrieval paradigm centered on resource organization, document positioning, and source traceability. In contrast, AI retrieval is analyzed as a compound process that integrates natural language understanding, query rewriting, task decomposition, multi-source recall, semantic matching, evidence filtering, contextual integration, reasoning, and generative output. DeepSeek's "deep thinking" and web search functions are used to illustrate the typical features of web-connected reasoning-based AI retrieval. In this process, web search provides external evidence, while deep thinking supports question decomposition, evidence selection, reasoning organization, and answer generation. This case-based comparison helps reveal that AI retrieval is not a simple extension of search engines, but a new retrieval form that combines retrieval, reasoning, and generation. [Results/Conclusions] The study found that traditional information retrieval still has irreplaceable advantages in precise search, authoritative verification, academic citation, legal and policy inquiry, and other scenarios requiring clear evidence chains and source traceability. Its strengths lie in certainty, interpretability, stable system performance, low operating cost, and direct access to original documents. However, traditional retrieval usually returns document lists, abstracts, or information fragments, which means that users still need to complete further reading, screening, comparison, and knowledge integration. AI retrieval, by contrast, shows stronger advantages in complex question answering, exploratory learning, interdisciplinary inquiry, research assistance, and knowledge consultation. Supported by semantic understanding, natural language interaction, and generative integration, it can reduce users' cognitive burden and transform scattered information into more structured and usable knowledge. Nevertheless, AI retrieval also brings new risks, including hallucination, inaccurate citation, source confusion, over-generalization, privacy leakage, model bias, and user over-reliance. Therefore, the future development of information retrieval should not be understood as a simple replacement of traditional retrieval by AI retrieval. A more reasonable direction is to construct a collaborative retrieval framework in which traditional retrieval provides reliable evidence anchors and AI retrieval undertakes semantic translation, knowledge organization, interactive explanation, and preliminary synthesis. To improve the credibility and applicability of AI retrieval, future systems should strengthen source citation, evidence verification, uncertainty indication, human review in high-risk scenarios, and user information literacy education. Future research may further conduct user experiments, system evaluation, and scenario-based empirical studies to examine the effectiveness, reliability, and governance mechanisms of AI retrieval in different information service contexts.

Key words: information retrieval, AI retrieval, artificial intelligence, large language model, DeepSeek