From Answers to Intermediaries: LLM Wikis as Knowledge Organization Layers in AI-Mediated Information Access
Generative artificial intelligence is reshaping information access by moving search interfaces from ranked records toward synthesized answers. In academic library contexts, however, answer-centered retrieval-augmented generation leaves unresolved problems of provenance, exploratory discovery, knowledge organization, trust calibration, and institutional accountability. This conceptual article reframes the problem by introducing the LLM Wiki as an AI-generated, semi-persistent, interlinked knowledge organization layer that mediates between library collections and users. Rather than treating retrieval as a hidden pipeline for answer production, the LLM Wiki reorganizes retrieved records, metadata, abstracts, and full text where available into inspectable topic pages, claim-source maps, source clusters, concept links, revision histories, and provenance trails. Drawing on information retrieval, exploratory search, knowledge organization, algorithmic accountability, and human-AI intermediation, the article develops a theory-building framework that distinguishes answer-centered RAG from wiki-mediated AI search. The framework identifies four core functions of the LLM Wiki: representational scaffolding, exploratory navigation, provenance exposure, and human-AI knowledge maintenance. The article further proposes design principles and an evaluation agenda for responsible implementation in academic library search. Its central contribution is to shift the study of generative AI search from answer generation toward accountable knowledge intermediation, where AI-generated representations become objects of navigation, verification, correction, and institutional stewardship.
artificial intelligence, library search, information retrieval, knowledge organization, exploratory search, retrieval-augmented generation, algorithmic accountability, human-AI interaction
1 Introduction
Generative artificial intelligence is changing the unit of library search. In the familiar discovery interface, the primary unit of interaction is the ranked record: a bibliographic item, metadata display, abstract, holdings link, or full-text access point. In many retrieval-augmented generation (RAG) systems, the primary unit becomes the generated answer: a synthesized response grounded, at least in design, in retrieved documents (Lewis et al. 2020; Gao et al. 2023). This shift has obvious appeal for academic libraries, where AI adoption has already been discussed in relation to reference, discovery, metadata, access services, and professional futures (Wheatley and Hervieux 2019; Cox et al. 2019). A conversational system can translate vague questions into plausible topical language, summarize unfamiliar literatures, and reduce the cognitive cost of moving across catalogs, databases, institutional repositories, and web-scale indexes. Yet it also creates a strategic problem for information science: when search is redesigned around answer production, the structures that make library search accountable, inspectable, and educative can become less visible.
This article argues that the more consequential question is not whether large language models (LLMs) can be attached to library discovery systems, but what model of information access such systems enact. Current research on neural and LLM-based information retrieval has focused heavily on technical roles such as representation learning, query rewriting, retrieval, reranking, reading, and search-agent orchestration (Mitra and Craswell 2018; Zhu et al. 2025). RAG extends this trajectory by joining parametric language models to non-parametric stores of external knowledge, thereby improving performance on knowledge-intensive tasks and offering a partial response to hallucination, staleness, and provenance problems (Lewis et al. 2020; Gao et al. 2023; Ji et al. 2023). These developments matter for libraries, but they do not exhaust the design space. A library search system is not merely an answer engine. It is also an institutional apparatus for representing collections, exposing evidentiary relationships, supporting exploratory inquiry, and mediating trust.
The theoretical gap is that existing models of AI-mediated information retrieval conceptualize retrieval primarily as support for answer generation, ranking, or conversational response rather than as the construction of persistent intermediary knowledge representations. Consequently, information retrieval theory lacks a sufficiently explicit level of analysis for explaining how AI-generated organizational structures reshape exploratory search, provenance inspection, user understanding, and institutional knowledge mediation. Academic libraries provide a strategically important domain for developing this argument, but the contribution is broader: it concerns AI-mediated information access wherever generated representations begin to mediate relations among sources, claims, users, and accountable institutions.
I therefore propose the LLM Wiki as a conceptual alternative to the answer-centered RAG interface. An LLM Wiki is an AI-generated, semi-persistent, interlinked knowledge layer that transforms retrieved library records, abstracts, full text where available, metadata, and authority data into navigable topic pages, source-backed claims, entity pages, concept links, and provenance trails. It is not simply a chatbot with citations, nor is it merely a generated encyclopedia. Its defining feature is intermediation: the LLM Wiki stands between collections and users as an inspectable knowledge organization layer whose pages can be searched, browsed, verified, corrected, and governed.
1.1 From Answer Generation to Knowledge Intermediation
The central claim of this article is that RAG-based search and LLM Wiki search represent different models of AI-mediated information access. RAG typically treats retrieval as a pipeline for answer generation. The system retrieves passages, conditions a model on those passages, and produces a response. By contrast, an LLM Wiki treats retrieval as the basis for constructing an intermediary information structure. The output is not only an answer, but a browsable arrangement of topics, claims, sources, and relationships.
This distinction is especially important in academic library contexts because scholarly search is often exploratory rather than merely transactional. Users frequently begin with unstable vocabularies, partial topic understanding, and evolving research questions. Classic information science models already treat uncertainty, iteration, and evolving need as central features of information seeking. Belkin (1980) framed information need through anomalous states of knowledge, while Kuhlthau (1991) described search as a process in which uncertainty, formulation, and learning unfold over time. Bates (1989) likewise challenged the idea of search as a single query matched against a stable target set, emphasizing instead iterative movement through documents, references, terms, and leads. Marchionini (2006) later distinguished lookup from exploratory search, where users move from finding toward learning, investigation, and understanding.
Seen against this lineage, answer-centered AI search risks solving too narrow a problem. A generated answer may satisfy a lookup need, but it can also flatten the search process by hiding alternative paths, suppressing uncertainty, and compressing source plurality into a single fluent response. The LLM Wiki reframes the design objective. Instead of asking only whether the system can answer a question, it asks whether the system can make a domain navigable, expose the evidentiary basis of generated claims, and leave behind a maintainable structure for further inquiry.
| Model | Primary output | Dominant interaction | Information science problem |
|---|---|---|---|
| Traditional discovery interface | Ranked records, facets, metadata displays, and links to holdings or full text. | Query, filter, inspect, and select. | How should collections be represented and retrieved? |
| RAG chatbot | Generated answer conditioned on retrieved passages or records. | Ask, receive, and follow up conversationally. | How should retrieved evidence be transformed into useful generated responses? |
| LLM Wiki | Interlinked topic pages, source-backed claims, related concepts, and provenance trails. | Search, browse, verify, refine, annotate, and govern. | How should AI-generated knowledge structures mediate discovery, interpretation, accountability, and maintenance? |
Table 1 positions the LLM Wiki as neither a replacement for traditional discovery nor a simple extension of RAG. Its difference lies in the status of the intermediate representation. In conventional discovery, the system represents collections primarily through metadata, indexing, ranking, and facets. In RAG chatbots, the intermediate retrieval process is usually operationally central but experientially minimized: users see the answer, sometimes with citations, but not always the structure of retrieval, selection, interpretation, and uncertainty that produced it. In an LLM Wiki, by contrast, the intermediate layer becomes the interface. The page, link, claim, source trail, and revision history become first-class information objects.
1.2 Why a Wiki Layer Matters for Libraries
The wiki metaphor is strategically useful because it foregrounds persistence, linkage, revisability, and social maintenance. These properties resonate with long-standing library and information science commitments. Knowledge organization systems are designed to organize materials for retrieval, collection management, and user access (Hodge 2000). They include classification schemes, subject headings, authority files, taxonomies, thesauri, and ontologies. The LLM Wiki does not replace these systems. Rather, it can be conceptualized as a machine-generated, user-facing knowledge organization layer that draws on them while also introducing new risks of error, bias, opacity, and over-synthesis, including risks already visible in subject representation and classification practice (Howard and Knowlton 2018).
This framing also clarifies why libraries need more than generic chatbot deployment. Academic libraries steward collections in which provenance, context, and interpretive plurality matter. A search interface that produces a polished answer without making its evidentiary and organizational work inspectable can miscalibrate trust. Generative AI for information access therefore raises sociotechnical questions about evaluation, harms, provenance, mitigation, and public knowledge infrastructures (Mitra et al. 2024). In library settings, those questions become institutional questions: Who may correct an AI-generated topic page? How are source-backed claims distinguished from model inferences? What happens when generated pages reproduce gaps in metadata, collections, or disciplinary vocabularies? What review responsibilities fall to librarians, faculty, students, vendors, and system designers?
The LLM Wiki answers these questions conceptually by relocating attention from the answer to the intermediary layer. It treats AI-mediated library search as a problem of accountable knowledge organization: the proposed model inserts an LLM Wiki layer between library collections and user interaction, then connects that layer to a governance loop through review, correction, versioning, and metadata feedback. This design commitment is often left implicit in AI search prototypes: the interface should not terminate at response generation. It should support a cycle in which generated structures can be inspected and improved. This cycle is consistent with emerging accounts of information and library professionals’ responsibilities in AI-augmented environments, especially responsibilities concerning inputs, uses, outputs, evaluation, education, and reflection (Marchionini 2024). It also aligns with broader work on algorithmic accountability, which cautions that transparency alone is insufficient unless connected to institutions, practices, and mechanisms of responsibility (Ananny and Crawford 2018; Shah 2018; Raji et al. 2020).
1.3 Contribution and Scope
This article makes three contributions to information science. First, it extends theories of AI-mediated information retrieval by introducing intermediary knowledge organization as a missing level of analysis between retrieval pipelines and user-facing answers. Second, it distinguishes answer-oriented RAG search from wiki-mediated LLM search as two different models of information mediation. Third, it conceptualizes the LLM Wiki as an intermediary knowledge organization layer that supports representational scaffolding, exploratory navigation, provenance exposure, and human-AI knowledge maintenance, then develops design principles and an evaluation agenda for responsible implementation.
The argument is conceptual and theory-building. It does not claim that LLM Wikis are inherently superior to RAG chatbots, traditional discovery layers, knowledge graphs, or librarian-authored subject guides. Rather, it claims that LLM Wikis reorganize the design problem. They shift attention from whether an AI system can produce a useful answer to whether an AI-mediated information system can create and maintain inspectable knowledge structures that support inquiry, verification, and institutional accountability. This shift is necessary because scholarly search is not only a matter of retrieving relevant items. It is also a matter of learning how a topic is organized, how claims are supported, where disagreement or uncertainty remains, and how users can responsibly move from sources to understanding.
2 Conceptual Approach and Guiding Questions
This article adopts a theory-building conceptual research design. Following guidance for conceptual articles, the aim is not to report empirical findings or review a literature exhaustively, but to synthesize fragmented theoretical conversations and construct a framework that explains a newly salient phenomenon (Jaakkola 2020; MacInnis 2011). The phenomenon is the emergence of AI-generated intermediary knowledge structures: semi-persistent pages, links, claims, source trails, and revision mechanisms that sit between retrieved sources and user understanding. This focus extends information science work on mediated and interactive retrieval by treating the intermediary not only as a person, interface, or dialogue process, but as a generated and governable representational layer (Wu and Liu 2003; Ruthven 2008). A theoretical contribution should specify the relevant constructs, explain relationships among them, and clarify why the proposed model changes how a field understands a phenomenon (Whetten 1989).
2.1 Conceptual Method
The conceptual method combines three moves. First, it synthesizes literatures on information retrieval, exploratory search, knowledge organization, algorithmic accountability, and human-AI interaction to identify a gap in current accounts of AI-mediated search. Second, it uses comparative analysis to distinguish the LLM Wiki from adjacent forms: traditional discovery interfaces, RAG chatbots, knowledge graphs, subject guides, and conventional wikis. Third, it constructs a theoretical model that specifies constructs, mechanisms, boundary conditions, design principles, and evaluation dimensions.
This method treats academic libraries as a theory-building domain rather than as the sole object of the contribution. Libraries are useful because they make the stakes of representation, provenance, authority, and governance unusually visible. However, the framework can also apply to other information environments in which AI-generated structures mediate access to documents, claims, policies, data, or scholarly records.
The validity of the conceptual argument depends on four criteria. Construct identification is warranted when a concept explains a distinct role in the mediation process rather than renaming an existing interface feature. Comparison is warranted when models are evaluated against shared dimensions such as representation, persistence, provenance, navigation, user agency, and governance. Abstraction is warranted when a concept can travel beyond the immediate library domain while preserving its explanatory role. Boundary specification is warranted when the framework identifies conditions under which the theory should not be expected to apply.
2.2 Limits of Existing Theoretical Traditions
The article does not claim that prior theories are wrong. Rather, it argues that each explains only part of AI-generated intermediary knowledge organization. Interactive information retrieval explains the coupling of users, systems, tasks, and representations, but it does not by itself theorize semi-persistent AI-generated knowledge objects as institutional artifacts (Ruthven 2008). Knowledge organization explains representation, warrant, classification, and subject relationships, but it has not yet fully accounted for LLM-generated pages and links that are produced dynamically from retrieval pipelines. Conversational search explains dialogue, clarification, and response generation, but it tends to foreground session-based interaction rather than durable representational layers (Zamani et al. 2023). Algorithmic accountability explains audit, transparency, and responsibility, but it does not specify how generated knowledge objects reshape exploratory navigation and user understanding.
The LLM Wiki reconciles these partial explanations by treating AI-mediated information access as a sequence of mediation transformations: retrieval becomes representation, representation becomes navigation, navigation supports understanding, and maintained representations become objects of institutional accountability. This is the theoretical architecture developed in the remainder of the article.
2.3 Guiding Theoretical Questions
The article is organized around four guiding theoretical questions.
| Question | Guiding theoretical question | Role in the argument |
|---|---|---|
| RQ1 | How should AI-generated intermediary knowledge structures be conceptualized within information retrieval theory? | Establishes the missing level of analysis. |
| RQ2 | In what ways do LLM Wikis differ theoretically from RAG systems as models of information mediation? | Clarifies the distinction between answer generation and knowledge intermediation. |
| RQ3 | How does an intermediary knowledge layer alter relationships among retrieval, organization, navigation, provenance, trust, and accountability? | Explains the mechanism through which generated representations affect search and understanding. |
| RQ4 | What principles should govern the design and evaluation of AI-generated knowledge organization layers? | Translates the theoretical framework into design and evaluation commitments. |
Table 2 clarifies the article’s theoretical trajectory. The first question defines the phenomenon. The second distinguishes it from dominant AI retrieval models. The third explains how it changes relationships among established information science constructs. The fourth translates the conceptual account into design and evaluation implications.
2.4 Formal Constructs
The framework rests on six constructs. These constructs are analytically distinct even though they are intertwined in working systems.
| Construct | Definition | Theoretical role |
|---|---|---|
| AI retrieval layer | The computational processes that retrieve, rank, extract, and condition generated outputs on sources. | Explains how sources enter the AI system. |
| Knowledge representation layer | The semi-persistent organizational structure through which retrieved materials are transformed into pages, links, labels, claims, and trails. | Introduces the missing level between retrieval and answers. |
| Intermediary knowledge objects | Inspectable objects such as topic pages, claim-source maps, source clusters, entity pages, relation links, and revision records. | Specifies what users and institutions can inspect and maintain. |
| Exploratory navigation | The user's movement through concepts, sources, relations, and uncertainties while refining an information need. | Links representation to information behavior. |
| User understanding | The user's developing grasp of a topic's structure, evidence base, alternative framings, and unresolved tensions. | Links navigation to learning and sensemaking. |
| Institutional accountability | The capacity of institutions and professionals to review, correct, version, audit, and govern generated representations. | Links AI search to stewardship, responsibility, and governance. |
Table 3 addresses a central conceptual problem: terms such as “wiki,” “intermediary,” “representation,” and “knowledge organization layer” can otherwise blur together. In this article, the retrieval layer is not the representation layer; the representation layer contains intermediary knowledge objects; those objects enable exploratory navigation; navigation supports user understanding; and institutional accountability governs the representation layer over time.
2.5 Explanatory Mechanism
The proposed mechanism is not simply that wiki-like interfaces are easier to browse. The theoretical claim is that AI-generated representations externalize latent relationships among sources, stabilize them as inspectable intermediary objects, and thereby change how users and institutions interact with retrieved information.
| Step | Mechanism | Expected theoretical effect |
|---|---|---|
| 1. Retrieval and extraction | The system retrieves records, passages, metadata, entities, concepts, and candidate claims. | Retrieval becomes the input to representation rather than only answer generation. |
| 2. Representation externalization | The LLM transforms latent semantic and evidentiary relationships into explicit pages, labels, links, and summaries. | Invisible retrieval and synthesis choices become visible enough to inspect. |
| 3. Intermediary object formation | Generated structures become semi-persistent objects that can be revisited, cited, compared, corrected, and versioned. | The interface gains durable objects of inquiry and governance. |
| 4. Navigational scaffolding | Users move through linked representations rather than reconstructing the topic space from result lists or answer text alone. | Cognitive reconstruction effort is reduced while exploratory paths remain open. |
| 5. Provenance inspection | Claim-source mappings and source trails allow users to inspect evidentiary support and calibrate trust. | Trust shifts from fluent acceptance toward evidence-based calibration. |
| 6. Knowledge maintenance | Librarians and other authorized actors revise, audit, suppress, approve, or enrich generated structures. | AI-mediated search becomes accountable knowledge mediation rather than disposable response production. |
Table 4 specifies why the LLM Wiki is theoretically different from an answer generator. It does not merely present more information. It changes the unit of mediation from the response to a set of intermediary objects that can support exploration, verification, and institutional governance.
2.6 Theoretical Propositions
The theory of AI-mediated knowledge intermediation can be stated as five propositions. These propositions are not empirical findings. They are theory-building claims that specify expected relationships among constructs and identify what later studies should test.
| Proposition | Theory-building claim | Implication for future research |
|---|---|---|
| P1 | When retrieved sources are transformed into semi-persistent intermediary knowledge objects, users gain more opportunities for exploratory navigation than when retrieved sources are transformed only into transient answers. | Compare exploratory movement across LLM Wikis, RAG chatbots, and discovery interfaces. |
| P2 | Intermediary knowledge objects improve user understanding by externalizing latent relationships among sources, concepts, claims, and evidence, thereby reducing the cognitive work required to reconstruct a topic space from result lists or answer text alone. | Measure topic learning, relation recognition, and query reformulation. |
| P3 | Claim-level provenance supports calibrated trust more strongly than answer-level citation display because it separates evidence, synthesis, inference, and review status. | Test whether users distinguish supported claims from unsupported or inferred claims. |
| P4 | Governable representations increase institutional accountability by making AI-generated organization separable from response generation and therefore available for review, correction, versioning, and audit. | Study librarian review workflows, correction propagation, and audit outcomes. |
| P5 | The value of an LLM Wiki depends on alignment among retrieval quality, representational warrant, exploratory usefulness, provenance inspectability, and maintenance capacity. | Evaluate boundary conditions, including sparse source bases, weak governance capacity, and contested categories. |
Table 5 elevates the framework from a descriptive model to a theory-building account. The propositions specify why intermediary knowledge objects matter: they persist, externalize relationships, expose provenance, enable governance, and depend on alignment across technical and institutional conditions.
3 Existing Models of AI-Mediated Retrieval and the Limits of Answer Generation
The rise of RAG and conversational search has made answer generation an increasingly natural design goal for information access systems. This is understandable. A user who asks a question often wants a direct response, not another list of links. Conversational information seeking research has therefore examined how systems can support mixed-initiative interaction, clarification, context maintenance, and dialogue-based search (Zamani et al. 2023). This trajectory builds on longer interactive and neural IR traditions in which representation, ranking, and user interaction are treated as coupled problems rather than isolated retrieval steps (Ruthven 2008; Mitra and Craswell 2018). Generative information retrieval likewise imagines systems that move beyond ranking documents toward directly generating identifiers, passages, responses, or structured representations of information needs (Metzler et al. 2021; Cao et al. 2024). These developments have expanded the technical vocabulary of search. They have also made it easier to mistake fluent response production for the whole of information access.
For libraries, that mistake is consequential. Academic library search supports known-item lookup, subject exploration, citation chasing, literature mapping, methodological learning, collection discovery, and credibility assessment. Some of these activities can be helped by generated answers, but few are exhausted by them. Library users often need to see how a search space is structured: which terms belong to a field, which authors or sources recur, which subtopics are adjacent, which claims are contested, and which records deserve closer inspection. A generated answer can summarize these relations, but a summary is not the same as an information environment in which users can inspect, navigate, and revise relations among sources.
The theoretical comparison in Table 6 uses nine dimensions to distinguish major models of information mediation. The comparison does not imply a simple historical progression or a hierarchy. Rather, it identifies what each model makes analytically visible and what each tends to leave undertheorized.
| Dimension | Traditional discovery | RAG chatbot | Knowledge graph | LLM Wiki |
|---|---|---|---|---|
| Primary retrieval unit | Record or result set | Retrieved passage or chunk | Entity and relation | Page, claim, source cluster, and relation |
| Dominant representation | Metadata, facets, ranking, holdings | Conversational generated answer | Formal graph structure | Interlinked generated knowledge objects |
| Persistence | Stable records and indexes | Often session-bound | Persistent graph objects | Semi-persistent and versioned |
| Provenance | Bibliographic and holdings metadata | Citation list or source links | Data lineage and ontology mappings | Claim-source maps and source trails |
| Navigation | Query, filter, inspect records | Dialogue and follow-up prompts | Entity traversal and semantic query | Search, browse, compare, verify, annotate |
| User agency | Select, refine, export, request access | Ask, clarify, regenerate, follow citations | Explore entities and relations | Explorer, verifier, annotator, corrector |
| Knowledge evolution | Metadata and collection updates | Model, index, or prompt updates | Graph maintenance and ontology revision | Revision, review, metadata feedback |
| Governance | Cataloging, metadata, access policies | System and vendor configuration | Ontology, data, and curation governance | Librarian review, audit, authority control |
| Explainability | Limited to records, facets, and ranking cues | Usually answer-level explanations | Relation paths and schema logic | Layered explanations of claims, links, pages, and review status |
Table 6 strengthens the article’s comparative method. Traditional discovery makes records and metadata visible; RAG makes answer synthesis visible; knowledge graphs make formal relations visible; LLM Wikis make generated intermediary knowledge objects visible. This is the conceptual basis for treating LLM Wikis as a new level of analysis rather than merely a different interface style.
The same comparison can be expressed at a higher theoretical level as a sequence of information mediation models. This sequence clarifies that the LLM Wiki is not primarily another interface category. It is an instantiation of a broader shift toward AI-mediated knowledge intermediation.
| Mediation model | Primary object of mediation | Dominant question | Theoretical limitation if isolated |
|---|---|---|---|
| Record mediation | Bibliographic records, metadata, result lists, and access links. | Which records are relevant and accessible? | Weak support for synthesis and exploratory sensemaking. |
| Answer mediation | Generated responses synthesized from retrieved sources. | What answer can be generated from retrieved evidence? | Weak visibility into representation, provenance, and revision. |
| Knowledge mediation | Semi-persistent pages, source clusters, claim-source maps, concept links, and revision histories. | How is a topic represented, navigated, and verified? | Risk of ungoverned generated structure. |
| Institutional mediation | Reviewed, governed, versioned, and audited knowledge representations connected to institutional workflows. | Who can maintain, correct, authorize, and account for generated representations? | Risk of governance burden without user-facing exploratory value. |
Table 7 reframes the comparison from systems to mediation models. Record mediation corresponds to traditional discovery. Answer mediation corresponds to RAG and conversational search. Knowledge mediation corresponds to the LLM Wiki as a layer of intermediary knowledge objects. Institutional mediation extends knowledge mediation into review, authority, correction, and accountability. The theory advanced here concerns the transition from answer mediation to accountable knowledge mediation.
3.1 The Appeal of Answer-Centered Search
Answer-centered search is attractive because it reduces visible complexity. Traditional discovery systems require users to translate an information need into keywords, interpret facets, compare records, assess metadata, and decide which sources merit attention. Decades of research on online catalogs and discovery interfaces show that these tasks can be difficult, especially when users lack domain vocabulary or when system vocabularies do not match user language (Borgman 1996; Asher et al. 2013). LLM interfaces seem to soften this mismatch. They can accept ordinary language queries, generate topical overviews, explain terminology, and produce a plausible first synthesis. In this sense, answer generation can function as a low-friction entry point into scholarly search.
The value of this entry point should not be dismissed. In exploratory contexts, users may benefit from conversational scaffolding that helps them articulate questions, identify synonyms, and move from vague interests toward searchable concepts. RAG systems can also reduce the epistemic isolation of purely parametric models by grounding responses in retrieved material (Lewis et al. 2020; Gao et al. 2023). For library search, this means that answer-centered systems may be useful for orientation, query expansion, preliminary synthesis, and reference-style assistance.
The problem is not that generated answers are useless. The problem is that they make one mode of information access unusually dominant. When the interface centers the answer, users may be encouraged to treat search as a question-response transaction rather than as an iterative process of inquiry, comparison, verification, and learning. This is a poor fit for academic work, where the user often needs not only an answer but also an account of how that answer was assembled.
3.2 Four Limits of the Answer Model
The limits of answer-centered search can be grouped into four information-science problems: compression, opacity, provenance fragility, and weak maintenance. These are not merely technical defects. They are structural tensions that emerge when retrieval, representation, and evaluation are subordinated to response generation.
| Limit | How it appears in answer-centered search | Why it matters for libraries | LLM Wiki response |
|---|---|---|---|
| Compression | Multiple records, claims, disagreements, and contextual distinctions are compressed into a single fluent response. | Users may lose sight of disciplinary plurality, minority positions, and alternative search paths. | Preserve topic pages, related concepts, source clusters, and unresolved tensions as navigable structures. |
| Opacity | Retrieval, ranking, passage selection, synthesis, and uncertainty handling are difficult for users to inspect. | Users cannot easily evaluate why particular sources, terms, or interpretations were foregrounded. | Expose the intermediary layer through pages, links, source trails, uncertainty notes, and revision history. |
| Provenance fragility | Citations may be absent, incomplete, mismatched, or too coarse to support claim-level verification. | Scholarly use requires traceability from claims to records, passages, metadata, and interpretive context. | Make claims source-addressable and distinguish cited evidence from model interpretation. |
| Weak maintenance | Generated responses are often session-bound and do not become durable objects for review, correction, or governance. | Libraries require accountable workflows for improving representations, correcting errors, and stewarding shared knowledge structures. | Treat generated pages as semi-persistent objects subject to annotation, review, versioning, and metadata feedback. |
Table 8 summarizes the central problem this article addresses. The first limit is compression. Generated answers are useful because they compress information, but compression can also remove the very structure that scholarly users need. A literature may contain multiple schools of thought, contested definitions, competing methods, and uneven representation across communities. An answer can mention these differences, but a wiki layer can make them navigable through linked topic pages, source clusters, and explicit notes about disagreement or uncertainty.
The second limit is opacity. In RAG systems, retrieval may be technically present but experientially hidden. Users may receive citations or source links, yet still have little access to why sources were selected, how passages were weighted, which alternatives were excluded, or what level of uncertainty attaches to the generated synthesis. This limitation matters because transparency in algorithmic systems is not merely a matter of showing more information. It requires usable mechanisms through which people can connect outputs to institutional practices of responsibility, review, and correction (Ananny and Crawford 2018; Raji et al. 2020).
The third limit is provenance fragility. RAG can improve grounding, but grounding does not automatically produce reliable scholarly citation behavior. Work on hallucination shows that language models can produce unsupported or distorted content even when outputs appear fluent and confident (Ji et al. 2023). In library search, the standard must be stricter than general plausibility. Users need to know whether a claim is directly supported by a source, inferred from multiple sources, derived from metadata, or generated as a system-level interpretation. An LLM Wiki can operationalize this distinction by attaching claims to records, passages, metadata fields, and confidence or uncertainty notes.
The fourth limit is weak maintenance. A generated answer is often ephemeral. It appears in a session, helps or misleads a user, and then disappears into logs or conversational history. Libraries, however, are maintenance institutions. They steward collections, vocabularies, metadata, authority records, guides, and instructional resources over time. If LLM outputs mediate access to collections, then those outputs should be objects of maintenance. The LLM Wiki makes this requirement visible by treating generated topic pages and claim structures as revisable, governable, and auditable information objects.
3.3 Why Existing Library Interfaces Do Not Fully Solve the Problem
Traditional discovery systems already provide some forms of structure: facets, subject headings, authority records, relevance ranking, citation links, availability indicators, and metadata displays. These are powerful and should remain central to library search. However, discovery systems often expose structure at the level of records and metadata rather than at the level of generated conceptual synthesis. They can show that items share a subject heading or author, but they do not usually construct an explanatory page that maps a topic, summarizes source clusters, identifies related concepts, and links claims to evidence. Conversely, RAG chatbots can construct explanations, but often without making the resulting conceptual structure persistent, inspectable, or maintainable.
Subject guides and librarian-authored research guides address a different part of the problem. They provide curated pathways into fields, databases, methods, and source types. Their strength is expert selection and pedagogical framing. Their limitation is scale and update burden: guides cannot be generated for every emerging topic, query formulation, interdisciplinary connection, or collection-specific cluster. The LLM Wiki should therefore not be framed as replacing discovery systems or subject guides. Its more plausible role is to occupy the space between them: dynamically generating provisional knowledge structures from collections while remaining accountable to library metadata, librarian review, and user correction.
Knowledge graphs also offer an important comparison. They represent entities and relations in formal structures that can support semantic search, recommendation, and reasoning. An LLM Wiki may draw on knowledge graph techniques, but it differs in rhetorical and interactional form. Its pages are designed for human interpretation as much as machine reasoning. They combine claims, narrative explanations, citations, links, and revision mechanisms. This human-facing quality matters because academic search is not only a problem of matching entities. It is also a problem of sensemaking.
3.4 Reframing the Design Problem
The limits of answer generation point toward a different design question. Instead of asking, “How can a library search system generate better answers?”, the LLM Wiki asks, “How can a library search system construct and maintain better intermediary knowledge structures?” The shift is subtle but fundamental. It changes the object of design from a response to an environment, from a session to a revisable layer, and from citation display to provenance governance.
This reframing does not make answer generation irrelevant. Answers remain useful as entry points, summaries, and navigational aids. But in an academic library setting, the answer should be treated as one view of a broader intermediary structure, not as the terminal product of search. The next section therefore defines the LLM Wiki more precisely and distinguishes it from adjacent forms: RAG chatbots, discovery layers, subject guides, knowledge graphs, and conventional wikis.
4 Defining the LLM Wiki
This section defines the LLM Wiki as a distinct information access form. The term is deliberately hybrid. LLM signals that the structure is produced or maintained in part through generative models that can summarize, label, link, and explain retrieved materials. Wiki signals that the structure is page-based, interlinked, revisable, and oriented toward shared knowledge maintenance rather than one-time response delivery. The combined term therefore identifies neither a generic chatbot nor a conventional encyclopedia. It names a library search intermediary: a semi-persistent knowledge organization layer generated from library collections and governed through human-AI maintenance.
For the purposes of this article, an LLM Wiki is defined as follows:
An LLM Wiki is an AI-generated, semi-persistent, interlinked knowledge organization layer that transforms retrieved library materials into navigable pages, source-addressable claims, concept links, entity pages, and provenance trails, with mechanisms for review, correction, and institutional governance.
Four features are essential to this definition. First, the LLM Wiki is collection-grounded. It is generated from library records, abstracts, full text where license permits, authority data, subject vocabularies, and other curated or locally relevant sources. Second, it is page-based and interlinked. It represents topics, authors, works, methods, events, places, and concepts as pages connected through explicit relations. Third, it is source-addressable. It should allow users to move from a generated claim to the records, passages, metadata fields, or authority sources that support it. Fourth, it is maintainable. Its pages are not disposable answers. They are revisable information objects that can be annotated, corrected, versioned, suppressed, approved, or enriched.
4.1 Core Components
The LLM Wiki has six core components. The first is a retrieval base, consisting of the records, documents, metadata, and authority data from which pages are generated. The second is a page generator, which creates provisional topic, entity, method, and source pages. The third is a linking layer, which identifies relationships among pages, including broader, narrower, related, cited-by, authored-by, and methodologically similar relations. The fourth is a claim-provenance layer, which maps claims to supporting sources and distinguishes direct evidence from synthesis or inference. The fifth is an interaction layer, through which users search, browse, compare, verify, and annotate pages. The sixth is a governance layer, through which librarians and other authorized actors review, correct, version, and maintain the wiki.
This architecture extends but does not collapse into existing models of information access. In RAG, retrieved evidence is often used to condition a generated answer (Lewis et al. 2020; Gao et al. 2023). In an LLM Wiki, retrieved evidence is used to construct a navigable structure. In knowledge graphs, entities and relations are formally represented for machine processing, semantic retrieval, and reasoning (Hogan et al. 2021). In an LLM Wiki, relations matter, but the interface is also rhetorical and pedagogical: it must help human users understand a topic, not merely encode relations. In subject guides, librarians curate pathways to resources for disciplines, assignments, or research tasks (Reeb and Gibbons 2004; Barker and Hoffman 2021). In an LLM Wiki, pages may be generated dynamically from collections, but their legitimacy depends on review, provenance, and alignment with library governance.
| Form | Primary unit | Main strength | Main limitation for library AI search |
|---|---|---|---|
| RAG chatbot | Generated answer | Low-friction question answering and synthesis | Often hides retrieval structure behind a fluent response |
| Library discovery layer | Bibliographic record or result set | Metadata-based retrieval, filtering, and access management | Usually weak at explanatory synthesis across records |
| Subject or research guide | Curated guide page | Expert curation and instructional framing | Difficult to scale across emerging or highly specific topics |
| Knowledge graph | Entity-relation triple or graph node | Formal relation modeling and machine-actionable semantics | May be less legible to ordinary users without narrative interfaces |
| Conventional wiki | Collaboratively edited page | Distributed editing and interlinked knowledge maintenance | May lack collection grounding, claim-level provenance, or professional review |
| LLM Wiki | Generated, source-addressable, revisable topic or entity page | Inspectable AI-mediated synthesis connected to sources, links, and governance | Requires careful design to prevent error propagation, overtrust, and maintenance burden |
Table 9 clarifies the boundary of the concept. The LLM Wiki borrows from each adjacent form, but its contribution comes from the combination. From RAG it borrows collection-grounded generation. From discovery layers it borrows institutional attachment to library metadata and access systems. From subject guides it borrows pedagogical orientation and topical navigation. From knowledge graphs it borrows explicit relation modeling. From conventional wikis it borrows linkage, revisability, and shared maintenance. The LLM Wiki becomes analytically useful because it brings these properties together around a single information science problem: how to make AI-mediated search inspectable, navigable, and governable.
4.2 Pages, Claims, Links, and Trails
The basic unit of an LLM Wiki is the page. Pages may correspond to topics, people, works, methods, theories, places, events, datasets, genres, or collections. A page should not be treated as a final statement of knowledge. It is better understood as a structured entry point into a search space. A topic page might contain a short overview, key terms, related concepts, major source clusters, representative works, points of disagreement, relevant databases, and suggested search paths. An author page might connect authority data, publications, coauthors, subjects, institutional affiliations, and related topics. A method page might identify works that use or discuss a particular method and distinguish methodological definitions from empirical applications.
Claims are the second unit. A claim is a generated proposition that can be inspected. In an academic library context, claims should be source-addressable. This means that users should be able to trace a claim to a record, passage, metadata field, authority source, or a specified inferential relation across sources. Source addressability is stricter than attaching a bibliography to a generated paragraph. It asks whether a user can determine which source supports which claim and whether the claim is quotation, paraphrase, synthesis, classification, or inference. This distinction is central because hallucination and unsupported synthesis are not only technical risks; they are failures of scholarly accountability (Ji et al. 2023; Mitra et al. 2024).
Links are the third unit. Links are not merely navigation aids. They are knowledge organization decisions. A link from “information behavior” to “exploratory search” proposes a relation; a link from a method to a community or corpus may foreground one intellectual pathway while hiding another. Because links can shape inquiry, they should be explainable and contestable. Some links may derive from subject headings, authority files, citation relations, or metadata. Others may be generated from text similarity, entity extraction, or model inference. The interface should make these differences visible.
Trails are the fourth unit. A provenance trail records how a page, claim, or link came to exist. It may include source records, retrieved passages, generation time, model or prompt version, review status, correction history, and responsible actors. Provenance has long been important for data quality and trust, and it has become increasingly relevant for explainable and trustworthy AI systems (Simmhan et al. 2005; Werder et al. 2022; National Institute of Standards and Technology 2024). In an LLM Wiki, provenance trails transform transparency from a generic aspiration into a navigational feature. Users do not simply receive an explanation; they can follow the construction of a page back toward the collection.
4.3 Semi-Persistence and Versioning
The phrase semi-persistent is important. An LLM Wiki page should persist long enough to be cited, revisited, corrected, and improved, but not so rigidly that it freezes generated interpretations as institutional fact. Semi-persistence allows the system to occupy a middle ground between the ephemerality of chatbot responses and the authority of catalog records or controlled vocabularies. A page may be generated provisionally, marked as unreviewed, revised by a librarian, annotated by users, updated when new sources are added, and retired when it becomes misleading or obsolete.
Versioning is therefore not a technical afterthought. It is a condition of accountable intermediation. Each version of a page should preserve enough history to answer questions such as: What sources supported this claim at the time it was shown? Which model or workflow generated it? Who reviewed it? What changed after correction? Which metadata records were updated as a result? These questions matter because libraries are not only service points; they are institutions of memory and trust. If AI-generated pages mediate access to collections, their histories become part of the accountability structure of the search system.
4.4 The LLM Wiki as Boundary Object
The LLM Wiki can also be understood as a boundary object: a shared structure that different communities can use for different purposes while coordinating around a common representational form (Star 2010; Huvila et al. 2017). Students may use the page as an entry point into a topic. Researchers may use it to identify related literatures or verify source clusters. Librarians may use it to detect metadata gaps, improve guides, or identify emerging collection needs. System designers may use it as an interface for surfacing retrieval diagnostics, provenance, and feedback signals. This multiplicity is a strength, but it also creates governance challenges.
The boundary-object interpretation matters because it links user-facing knowledge mediation to institutional mediation. A topic page is not only a search aid; it is also a coordination object among users, librarians, metadata specialists, collection stewards, and system designers. Because the LLM Wiki sits between collections and users, it can redistribute authority. It may make hidden relations visible, but it may also amplify collection biases, normalize dominant terminology, or present inferred relations as more stable than they are. Human-centered AI scholarship emphasizes that effective systems require more than explanation; they require meaningful agency, contestability, and adaptation within real use contexts (Amershi et al. 2019; Shneiderman 2020). For the LLM Wiki, this means that users and librarians must be able to question, correct, and reshape the intermediary layer rather than simply consume it.
4.5 Boundary Conditions
The LLM Wiki concept is most applicable where four conditions hold. First, the information environment must contain a sufficiently rich source base: metadata, documents, abstracts, authority data, or other materials from which source-addressable pages and claims can be generated. Second, users must benefit from exploratory movement across concepts, sources, and relationships rather than only from direct lookup. Third, provenance and accountability must matter because generated representations may influence scholarly, institutional, professional, or public decisions. Fourth, there must be some governance capacity, whether through librarians, domain experts, community review, policy stewards, or other accountable actors.
The concept is less appropriate where information needs are purely transactional, where source bases are too sparse to support reliable representation, where generated pages would expose restricted or sensitive content, or where no institution can review and maintain the intermediary layer. It is also less appropriate when generated categories cannot satisfy adequate cultural, ethical, or community warrant for representation (Choi 2022). These boundaries are theoretically important. They prevent the LLM Wiki from becoming a universal solution and clarify the conditions under which intermediary knowledge organization is likely to add value.
4.6 Working Definition for the Remainder of the Article
The remainder of this article uses LLM Wiki to refer to a collection-grounded, semi-persistent, interlinked, source-addressable, and governable knowledge organization layer for AI-mediated information access, with academic library search as the primary domain of analysis. This working definition excludes three weaker uses of the term. It does not refer to any chatbot that cites sources. It does not refer to an automatically generated encyclopedia detached from local collections. It does not refer to a knowledge graph with a thin textual interface. The LLM Wiki is instead a hybrid information intermediary: generated enough to scale, structured enough to support exploration, cited enough to support verification, and governed enough to belong in a library setting.
5 Theoretical Foundations
The LLM Wiki is not only a technical interface pattern. It is a theoretical proposition about how generative AI should mediate information access in library settings. Its value depends on whether it can connect computational capabilities to established information science concerns: how people search under uncertainty, how knowledge is represented and organized, how systems make evidence visible, and how human judgment remains active in mediated inquiry. This section develops four theoretical foundations for the LLM Wiki: exploratory search and sensemaking, knowledge organization, algorithmic accountability, and human-AI intermediation.
These foundations matter because they prevent the LLM Wiki from being reduced to a novel interface metaphor. A top-tier information science contribution cannot simply claim that a generated wiki would be useful. It must explain why this form responds to enduring problems in information access. The LLM Wiki is therefore grounded in a claim about intermediary structure: users often need more than retrieved documents and more than generated answers. They need inspectable arrangements that help them move among concepts, claims, sources, and uncertainties while retaining the ability to question how those arrangements were produced.
The four lenses are not used as separate background literatures. They form an integrated explanatory chain. Exploratory search explains why users need navigable structures under uncertainty. Knowledge organization explains why generated pages and links are representational decisions. Algorithmic accountability explains why those decisions require provenance, review, and contestability. Human-AI intermediation explains how users, librarians, and systems share the work of maintaining the intermediary layer. Together, these lenses explain how AI-generated representations can become accountable knowledge mediation rather than disposable answer production.
| Lens | Core concern | Implication for LLM Wikis | Design risk if ignored |
|---|---|---|---|
| Exploratory search and sensemaking | How users move through uncertainty, learn a domain, reformulate questions, and construct meaning. | Pages should support browsing, comparison, reformulation, and topic learning rather than only answer delivery. | The system becomes a fluent lookup tool that weakly supports learning and research development. |
| Knowledge organization | How concepts, documents, entities, subjects, and relations are represented for access and interpretation. | Generated pages and links should be treated as knowledge organization decisions, not neutral summaries. | Generated structures naturalize biased, incomplete, or unstable categories without professional scrutiny. |
| Algorithmic accountability | How systems expose responsibility, provenance, limitations, review mechanisms, and institutional control. | Claims, links, and pages should include provenance, uncertainty, versioning, and review status. | Users overtrust outputs because the system offers citations without meaningful traceability or responsibility. |
| Human-AI intermediation | How people and AI systems coordinate work across generation, verification, correction, and maintenance. | Librarians and users should be able to contest, annotate, correct, and improve the intermediary layer. | AI output becomes detached from the human practices that make library information trustworthy. |
Table 10 summarizes the role of each lens in the conceptual argument. The table also identifies the risk created when any one lens is omitted. Without exploratory search, the LLM Wiki becomes a better answer box. Without knowledge organization, it becomes an unexamined set of generated labels and links. Without accountability, it becomes a persuasive but weakly governable synthesis machine. Without human-AI intermediation, it becomes a system that displaces rather than strengthens the professional and user practices through which libraries maintain trust.
5.1 Exploratory Search and Sensemaking
The first foundation is exploratory search. Library search often begins before a user knows exactly what they are looking for. Students may begin with a broad topic assigned in a course. Researchers may enter an unfamiliar adjacent field. Faculty may look for emerging literatures, methods, or debates that do not yet have stable terminology. In these cases, the user’s problem is not simply to retrieve a known item or maximize precision for a stable query. The problem is to learn enough about a domain to formulate better questions.
Exploratory search theory is useful because it treats searching as learning, investigation, and sensemaking rather than only retrieval (Marchionini 2006). It resonates with earlier models of information seeking that emphasize uncertainty, formulation, iteration, and changing information needs (Belkin 1980; Bates 1989; Kuhlthau 1991). Sensemaking approaches similarly describe users as constructing meaning across gaps, situations, and partial information rather than merely consuming retrieved facts (Dervin 1998; Pirolli and Card 2005).
The LLM Wiki supports exploratory search by externalizing intermediate structures. Instead of generating a single answer to a query, it can display a topic’s conceptual neighborhood: broader and narrower terms, related methods, major source clusters, recurring authors, contested definitions, and possible search paths. This matters because exploratory search is often path-dependent. The concepts a user sees early can shape the questions they ask later, and searching-as-learning research has emphasized the need for measures and interface conditions that capture learning, control, and transparency rather than only retrieval success (Rieh et al. 2014; Sciascio et al. 2020). A wiki layer can make those paths visible and revisable.
This lens also limits what should be claimed for the LLM Wiki. It should not be evaluated only by whether its first generated response is correct. It should be evaluated by whether it helps users learn a topic, discover useful terminology, identify relevant sources, compare perspectives, reformulate queries, and preserve awareness of uncertainty. In other words, exploratory value lies in the quality of the user’s movement through the information environment.
5.2 Knowledge Organization
The second foundation is knowledge organization. An LLM Wiki generates labels, links, summaries, topic boundaries, and entity relations. These are not merely interface conveniences. They are acts of organization. Knowledge organization research has long examined how subjects, concepts, documents, and relations are represented for retrieval and understanding (Hjørland 2008; Hodge 2000). It also cautions that organization is never neutral: classification, indexing, and subject representation reflect epistemic assumptions, institutional histories, disciplinary priorities, and social values.
This perspective is central to the LLM Wiki because generated pages can appear natural even when their categories are contingent. A model may group sources under a topic label that is familiar in one discipline but misleading in another. It may infer a relation between concepts because they co-occur in text, even when specialists would treat the relation as weak, historical, or contested. It may reproduce dominant vocabularies while underrepresenting Indigenous, local, non-English, or marginalized knowledge traditions. Studies of classification and subject headings show that representational bias is not hypothetical but embedded in long-standing access infrastructures (Howard and Knowlton 2018). These risks are knowledge organization risks, not only model-performance risks.
An LLM Wiki therefore needs explicit relationships to established library knowledge organization systems. Subject headings, classification schemes, authority files, thesauri, ontologies, and local metadata practices can function as anchors, constraints, and points of comparison. They should not be treated as perfect ground truth, since legacy systems also contain biases and exclusions. But they provide institutional memory and professional infrastructure against which generated organization can be evaluated.
From this lens, the LLM Wiki is best understood as a machine-generated knowledge organization interface. Its pages and links should be inspectable as provisional organizational decisions. The system should show when a relation is derived from an authority file, when it is inferred from retrieved text, when it is generated by a model, and when it has been reviewed by a librarian. The goal is not to eliminate machine-generated organization, but to make its status visible and contestable.
5.3 Algorithmic Accountability and Provenance
The third foundation is algorithmic accountability. Generative AI search systems can produce persuasive outputs that obscure the social and technical processes behind them. This is especially problematic in library settings, where users rely on information systems to support scholarly judgment. The issue is not only whether the model is transparent in some abstract sense. As Ananny and Crawford (2018) argue, transparency alone can be an insufficient ideal if it does not connect to practices of responsibility. Accountability requires mechanisms through which outputs can be questioned, reviewed, corrected, and governed.
For the LLM Wiki, accountability begins with provenance. A generated topic page should make clear which records, passages, metadata fields, authority sources, or inferred relations support its claims. Provenance trails can help users distinguish direct evidence from synthesis, paraphrase from inference, and source-backed claims from model-generated interpretation. This distinction is particularly important given the documented risks of hallucination and unsupported generation in natural language systems (Ji et al. 2023). It also aligns with emerging generative AI risk management guidance, which emphasizes mapping, measuring, managing, and governing AI risks across system lifecycles (National Institute of Standards and Technology 2024).
Accountability also requires institutional roles. If an LLM Wiki page is wrong, outdated, biased, or misleading, who can correct it? If a relation between concepts is disputed, how is that dispute recorded? If a generated page depends on licensed full text, what may be exposed to users? If a source is removed from a collection or metadata is corrected, how does the page change? These questions cannot be answered by model architecture alone. They require governance workflows, review status indicators, version histories, and policies for suppression, correction, and escalation.
The LLM Wiki therefore shifts accountability from the isolated output to the maintained intermediary layer. A chatbot answer can be evaluated after the fact, but a wiki page can be reviewed, revised, and situated in an institutional record. This does not remove risk. It makes risk more visible and more actionable.
5.4 Human-AI Intermediation
The fourth foundation is human-AI intermediation. Libraries have always relied on intermediaries: catalogers, reference librarians, indexers, subject specialists, system designers, vendors, and increasingly algorithmic ranking and recommendation systems. The LLM Wiki adds another intermediary, but it should not replace human judgment. It should reorganize the relationship between machine assistance and human stewardship.
Human-centered AI scholarship emphasizes that AI systems should support intelligibility, agency, appropriate trust, and human control across real use contexts (Amershi et al. 2019; Shneiderman 2020). For library search, this means that the user should not be positioned only as a recipient of generated text. The user should be able to inspect sources, follow links, compare alternatives, flag problems, and refine the structure. Librarians should not be positioned only as end users of a vendor system. They should have tools for reviewing pages, adjusting authority mappings, correcting claims, documenting local policies, and feeding improvements back into metadata and instruction.
This lens reframes intermediation as distributed work. The LLM may generate initial pages, identify candidate relations, and draft source summaries. Users may browse, verify, annotate, and expose gaps. Librarians may validate, correct, suppress, or enrich generated structures. System designers may monitor retrieval quality, provenance integrity, and error patterns. The LLM Wiki becomes the shared surface on which these forms of work meet.
Human-AI intermediation also introduces a design ethic: the system should make human judgment easier to exercise, not easier to bypass. A useful LLM Wiki should slow users down at the right moments by showing uncertainty, surfacing alternative sources, and distinguishing evidence from interpretation. It should also speed users up where speed is appropriate, such as identifying relevant terminology, clustering sources, or revealing connections across collections. The aim is calibrated mediation rather than frictionless automation.
5.5 Synthesis
Taken together, these foundations define the LLM Wiki as a theory of accountable exploratory intermediation. Exploratory search explains why users need navigable structures rather than only answers. Knowledge organization explains why generated pages and links must be treated as representational decisions. Algorithmic accountability explains why provenance, review, and versioning are not optional interface features. Human-AI intermediation explains why the system must support shared work among users, librarians, and computational agents.
This synthesis prepares the conceptual framework developed in the next section. If the LLM Wiki is an intermediary knowledge organization layer, then its functions can be specified more precisely. The following section identifies four such functions: representational scaffolding, exploratory navigation, provenance exposure, and human-AI knowledge maintenance.
6 The LLM Wiki Intermediary Framework
The preceding sections established the conceptual need for an intermediary layer between library collections and users. This section specifies that layer more directly. The LLM Wiki Intermediary Framework conceptualizes wiki-mediated AI search as four linked functions: representational scaffolding, exploratory navigation, provenance exposure, and human-AI knowledge maintenance. These functions describe what the LLM Wiki must do if it is to operate as an information science contribution rather than as a chatbot interface with a different visual form.
The framework is intentionally functional rather than platform-specific. It does not prescribe a particular model, database, vector store, metadata standard, ranking algorithm, or front-end implementation. Instead, it identifies the information work that a responsible LLM Wiki must support in academic library contexts. The same framework could guide a lightweight prototype built over abstracts and metadata, a local system connected to an institutional repository, or a more ambitious discovery interface integrated with licensed content and authority data.
Figure 1 visualizes the framework as a theory-building model of AI-mediated information access. The figure contrasts answer-centered RAG with the LLM Wiki model, then places the intermediary knowledge organization layer at the center because the article’s main argument is that generative AI search should not terminate in a generated answer. Instead, retrieved materials should be reorganized into inspectable, linked, and versioned knowledge objects that support navigation, verification, maintenance, and accountability. The surrounding retrieval, outcome, and feedback layers show that the LLM Wiki depends on both computational extraction and institutional stewardship.
| Function | Primary question | Typical interface objects | Evaluation concern |
|---|---|---|---|
| Representational scaffolding | How can retrieved materials be transformed into intelligible topic, entity, method, and source structures? | Topic pages, source clusters, entity pages, method pages, concept summaries, terminology notes. | Does the structure help users understand the shape and boundaries of a topic? |
| Exploratory navigation | How can users move across concepts, sources, disciplines, and uncertainty while developing a research question? | Related topics, broader and narrower concepts, source trails, alternative search paths, interdisciplinary links. | Does the interface support learning, reformulation, comparison, and discovery beyond the initial query? |
| Provenance exposure | How can users inspect the evidentiary basis, generation history, and uncertainty of pages, claims, and links? | Claim-source maps, cited passages, metadata traces, uncertainty labels, generation logs, review status. | Can users verify claims and calibrate trust without excessive cognitive burden? |
| Human-AI knowledge maintenance | How can users, librarians, and systems correct, version, govern, and improve the intermediary layer over time? | Annotations, correction queues, version histories, librarian review workflows, authority mappings, metadata feedback. | Can the institution maintain quality, correct errors, and prevent the generated layer from drifting away from collection realities? |
Table 11 presents the framework as a set of design and evaluation commitments. Each function addresses a different weakness of answer-centered AI search. Representational scaffolding responds to the compression of complex literatures into short answers. Exploratory navigation responds to the need for iterative inquiry rather than one-shot response delivery. Provenance exposure responds to opacity and citation fragility. Human-AI knowledge maintenance responds to the institutional problem of keeping generated structures accurate, accountable, and aligned with collection development, metadata, and user needs.
6.1 Representational Scaffolding
Representational scaffolding is the function through which the LLM Wiki converts dispersed records into intelligible structures. The term scaffolding is important because the wiki layer does not replace the user’s own inquiry. It supports that inquiry by making an unfamiliar domain easier to enter, compare, and revise. In exploratory search, scaffolding can help users develop search vocabulary, recognize topic boundaries, and identify promising pathways through complex information spaces (Marchionini 2006; Bates 1989). More recent work on search scaffolding and searching as learning similarly emphasizes support for learning-oriented search processes, control, and transparency rather than simple retrieval completion (Rieh et al. 2014; Sciascio et al. 2020; Vakkari 2016; Zhang et al. 2020).
In an LLM Wiki, representational scaffolding may appear as topic overviews, concept maps, entity pages, method pages, source clusters, terminology notes, and summaries of disciplinary perspectives. A student searching for “AI in libraries,” for example, may need to distinguish library automation, machine learning for metadata, recommendation systems, generative AI for reference, algorithmic bias in discovery, and AI literacy instruction. A traditional result list may contain sources relevant to all of these areas. A RAG answer may summarize some of them. A wiki layer can instead create a navigable structure that shows how they relate and where they diverge.
This function should be designed with restraint. The LLM Wiki should not imply that generated structures are definitive maps of a field. Representational scaffolding is provisional. It should expose alternative labels, contested categories, and gaps in the collection. Where possible, it should connect generated topics to established vocabularies, subject headings, classification terms, and authority data while also indicating when generated labels do not map cleanly onto existing systems. This makes the scaffold useful without pretending that the scaffold is the territory.
6.3 Provenance Exposure
Provenance exposure is the function through which the LLM Wiki makes the basis of its pages, claims, and links inspectable. In answer-centered RAG, citations are often presented as evidence that the answer is grounded. However, citation display alone does not guarantee that each claim is supported by the cited source, that the source was interpreted correctly, or that important retrieved sources were not omitted. Explainable information retrieval research emphasizes the need to make search systems more transparent and trustworthy by explaining rankings, relevance, and source selection (Anand et al. 2022). Provenance research likewise treats the history and derivation of information objects as central to trust and reuse (Simmhan et al. 2005).
For the LLM Wiki, provenance exposure should operate at multiple levels. At the page level, users should see which sources contributed to a topic page and when the page was last generated or reviewed. At the claim level, users should see which passage, record, or metadata field supports a statement. At the link level, users should see whether a relation came from an authority file, citation relation, subject heading, co-occurrence pattern, embedding similarity, or model inference. At the system level, users and librarians should see generation settings, review status, and known limitations.
The purpose of provenance exposure is trust calibration, not maximal disclosure for its own sake. Too much raw information can overwhelm users, while too little can create false confidence. The interface should therefore layer provenance. A casual user may need a visible source trail and review badge. A librarian may need retrieval logs, claim-source mappings, and correction histories. A researcher may need exportable citations and evidence snippets. The LLM Wiki should make these layers available according to task and expertise.
6.4 Human-AI Knowledge Maintenance
Human-AI knowledge maintenance is the function through which the LLM Wiki remains accountable over time. This function is the clearest difference between a wiki layer and an answer generator. If a generated answer is wrong, the user may ignore it, regenerate it, or report it. If a generated wiki page is wrong, the system needs a way to correct the page, preserve the correction, propagate the change, and prevent the same error from recurring. Maintenance therefore transforms AI search from a session-level interaction into an institutional workflow.
This function draws on the library’s long-standing role as a maintenance institution. Libraries maintain metadata, authority records, collection descriptions, subject guides, instructional materials, and access systems. AI-generated layers should be incorporated into that maintenance ecology rather than placed outside it. Recent work on AI and metadata management emphasizes the importance of benchmarking, maintenance responsibility, and human validation in AI-assisted metadata workflows (Program for Cooperative Cataloging Task Group on AI and Machine Learning for Metadata 2025). The same logic applies to LLM Wiki pages. Generated organization needs human review, especially where errors affect access, representation, or trust.
Human-AI knowledge maintenance may include annotation tools, correction queues, librarian approval states, version histories, model feedback mechanisms, page retirement policies, and metadata enrichment workflows. It may also include community reporting mechanisms, although these should be carefully governed to prevent abuse and preserve professional accountability. The goal is not to make every user a cataloger. The goal is to create a pathway through which observed errors and insights can improve the shared intermediary layer.
Maintenance should also include periodic auditing. Pages may drift as collections change, models update, vocabularies evolve, and new scholarship appears. A page that was useful at one point may become misleading later. Audits should therefore examine source coverage, claim support, bias and omission, link quality, review status, and user feedback. This makes the LLM Wiki not only a search interface but a site of ongoing collection intelligence.
6.5 Framework Logic
The four functions are mutually dependent. Representational scaffolding without provenance exposure can become polished but untrustworthy synthesis. Provenance exposure without exploratory navigation can become a burdensome audit interface. Exploratory navigation without knowledge maintenance can propagate outdated or biased relations. Human-AI knowledge maintenance without useful scaffolding may produce a governance burden with little user value. The framework therefore treats the LLM Wiki as an integrated intermediary system.
The sequence of functions also describes a workflow. The system first organizes retrieved materials into provisional structures. Users then navigate those structures as part of exploratory inquiry. As they inspect pages, they encounter provenance features that help them verify and calibrate trust. Their interactions, corrections, and librarian reviews then feed maintenance processes that improve the intermediary layer. In this way, the LLM Wiki becomes a dynamic knowledge organization environment rather than a static search result.
This framework provides the conceptual basis for the design principles developed in the next section. If LLM Wikis are to serve academic libraries, their interfaces must expose knowledge structures, make claims source-addressable, keep generated organization contestable, distinguish evidence from interpretation, preserve traditional retrieval pathways, and place library professionals in governance roles.
7 Design Principles for Wiki-Mediated Library AI Search
The LLM Wiki Intermediary Framework implies a set of design principles for AI-mediated library search. These principles are not interface preferences in a narrow usability sense. They are normative commitments about what library search should preserve when generative AI becomes part of discovery: source traceability, conceptual plurality, human judgment, institutional accountability, and support for learning. They also translate the framework into requirements that can guide prototype development, procurement, evaluation, and governance.
The principles below are intentionally stated at a level that can apply across different technical architectures. A local prototype may implement them through metadata-only pages and manually reviewed source trails. A production system may implement them through retrieval logs, claim-level passage links, authority integration, and librarian review dashboards. The central question is not whether a system uses a particular model or database. The question is whether the system makes AI-mediated search inspectable, contestable, and maintainable in ways that align with library values.
| Principle | Design requirement | Primary risk addressed |
|---|---|---|
| Expose knowledge structures, not only answers | Show topic pages, source clusters, related concepts, uncertainty, and alternative paths alongside summaries. | Over-compression of complex literatures into fluent but flattening answers. |
| Make every generated claim source-addressable | Connect each substantive claim to records, passages, metadata fields, authority sources, or stated inferential support. | Citation theater, unsupported synthesis, and weak scholarly verification. |
| Keep AI-generated organization contestable | Allow users and librarians to flag, annotate, correct, suppress, or revise generated pages and links. | Naturalization of biased, erroneous, outdated, or locally inappropriate generated categories. |
| Distinguish evidence from interpretation | Mark whether content is directly stated by a source, synthesized across sources, inferred by the model, or supplied by metadata. | Confusion between evidence, paraphrase, synthesis, and speculation. |
| Support both exploratory browsing and precise retrieval | Preserve known-item search, filtering, citation export, and record inspection while adding wiki-based navigation. | Replacement of established library search affordances with a brittle conversational interface. |
| Govern the intermediary layer through library stewardship | Provide review states, version histories, authority mappings, correction workflows, and audit mechanisms. | Drift of generated structures away from professional accountability and collection realities. |
Table 12 summarizes six design principles and the risks each principle addresses. Together, they specify how an LLM Wiki should behave as a library search intermediary. They also provide an early evaluation checklist: a proposed system that cannot expose knowledge structures, map claims to sources, support correction, distinguish evidence from inference, preserve precise retrieval, and sustain governance should not be treated as a mature library AI search system.
7.1 Principle 1: Expose Knowledge Structures, Not Only Answers
AI-mediated library search should reveal the structure of a topic rather than only produce a response to a query. A generated answer may be useful as an overview, but it should be embedded within a broader arrangement of topic pages, related concepts, source clusters, terminology notes, and search pathways. This principle follows from exploratory search and searching-as-learning theory, which treat search as learning, investigation, control, and transparency rather than only fact retrieval (Marchionini 2006; Vakkari 2016; Rieh et al. 2014; Sciascio et al. 2020).
In practice, this means that an LLM Wiki should show how a generated answer fits into a wider information environment. If a user asks about “algorithmic bias in library discovery,” the system should not merely summarize the issue. It should expose adjacent concepts such as ranking bias, metadata bias, subject heading bias, recommender systems, audit methods, and critical cataloging. It should also show relevant source clusters, recurring authors, methodological approaches, and contested terminology. The answer becomes one navigational object among others, not the endpoint of search.
This principle also protects against excessive compression. Academic topics often contain disagreement, uncertainty, and disciplinary variation. When these are compressed into one fluent response, users may lose sight of alternative interpretations. Exposing knowledge structures helps users see that a topic has shape, boundaries, and unresolved tensions.
7.2 Principle 2: Make Every Generated Claim Source-Addressable
The LLM Wiki should make substantive generated claims traceable to specific sources. Source addressability means more than attaching a bibliography to a page. It requires a claim-source relationship that users can inspect. A claim may be supported by a passage, a metadata field, an authority record, a citation relation, or a synthesis across multiple sources. The interface should make that support visible at an appropriate level of detail.
This principle responds to a central weakness of answer-centered RAG: citations may create an appearance of grounding without guaranteeing that the cited source supports the exact claim. Research on hallucination shows that generated text can be fluent and unsupported (Ji et al. 2023). Explainable AI guidance similarly emphasizes that systems should supply evidence, reasons, and limits in forms that users can understand and act upon (Phillips et al. 2021; Anand et al. 2022). For scholarly search, the requirement is especially strict because users may use generated claims in literature reviews, assignments, grant proposals, and research decisions.
Source-addressable claims should therefore support multiple levels of inspection. A novice user may click a claim to see a short evidence snippet and citation. A researcher may need page numbers, DOI links, metadata fields, and exportable references. A librarian may need retrieval logs and information about which source versions were used. The design challenge is to provide layered evidence without overwhelming users.
7.3 Principle 3: Keep AI-Generated Organization Contestable
Generated pages, labels, and links should be contestable. This principle follows from knowledge organization theory: categories and relations are not neutral containers but representational decisions (Hjørland 2008). When an LLM Wiki groups records into a topic page, labels a source cluster, or links two concepts, it is making a knowledge organization move. Such moves should be open to review and correction.
Contestability should be built into the interface. Users should be able to flag a misleading page, question a link, report a missing source, or suggest an alternative label. Librarians should be able to accept, reject, revise, suppress, or annotate generated structures. Review status should be visible, so that users can distinguish unreviewed generated pages from librarian-reviewed or institutionally curated pages.
Contestability is especially important where generated organization intersects with local, cultural, or disciplinary knowledge. A model may prefer dominant English-language terminology, flatten local histories, or reproduce collection biases. Library governance should therefore treat generated structures as provisional. The LLM Wiki should invite scrutiny rather than asking users to trust organization because it appears polished.
7.4 Principle 4: Distinguish Evidence From Interpretation
The LLM Wiki should clearly distinguish what sources say from what the system infers. This distinction is central to scholarly accountability. A source-backed claim, a multi-source synthesis, a model-generated interpretation, and a metadata-derived label are different kinds of statements. They should not be presented with the same epistemic status.
In practice, pages may use labels such as “source states,” “synthesized from,” “inferred relation,” “metadata-derived,” or “librarian-reviewed.” The exact vocabulary can vary, but the interface should prevent users from mistaking interpretation for evidence. This is consistent with provenance documentation work, which emphasizes that trustworthy AI systems require information about data origins, transformations, and derivations (Simmhan et al. 2005; Chari et al. 2023).
This principle also supports better user education. Academic library search is not merely about finding sources; it is about learning how evidence works. By distinguishing evidence from interpretation, the LLM Wiki can teach users to ask better questions about generated outputs. What is directly supported? What is inferred? What is uncertain? What should be checked in the source itself?
7.5 Principle 5: Support Both Exploratory Browsing and Precise Retrieval
The LLM Wiki should complement traditional library search rather than replace it. Users still need known-item lookup, advanced search, facets, filters, record inspection, availability information, citation export, and links to full text. A wiki layer should add exploratory and interpretive affordances without weakening precise retrieval.
This principle matters because conversational AI interfaces can make search feel simpler while hiding the controls that expert users need. A student may benefit from a topic page, but a researcher may need to inspect exact records, apply date limits, export citations, or search within a specific database. A librarian may need to troubleshoot metadata, access rights, or indexing behavior. The LLM Wiki should therefore be layered over, or integrated with, conventional retrieval affordances.
The design goal is mode switching. Users should be able to move from a generated topic page to a record, from a record to a source cluster, from a source cluster to a filtered result set, and from a result set back to a revised topic page. This supports both exploratory browsing and precise retrieval as complementary activities.
7.6 Principle 6: Govern the Intermediary Layer Through Library Stewardship
The LLM Wiki should be governed as part of the library’s information infrastructure. Governance includes review workflows, version histories, authority control, metadata feedback, audit procedures, bias and omission checks, licensing constraints, and policies for correction or suppression. This principle follows from the claim that the LLM Wiki is not a disposable output but a semi-persistent intermediary layer.
Governance should also assign roles. Librarians may review high-impact pages, approve authority mappings, monitor correction queues, and identify metadata improvements. Users may flag issues and contribute annotations. System administrators may monitor retrieval performance, model updates, and security. Vendors or developers may provide logging, export, audit, and configuration tools. Without role clarity, generated pages can become institutionally ambiguous: influential enough to shape search, but not clearly owned by anyone.
This principle is aligned with broader AI risk management and human-centered AI guidance, which emphasize governance, measurement, human oversight, and lifecycle management (National Institute of Standards and Technology 2024; Shneiderman 2020). It is also aligned with library metadata governance, where automation must be paired with validation, documentation, and professional responsibility (Program for Cooperative Cataloging Task Group on AI and Machine Learning for Metadata 2025).
7.7 Design Implications
These principles imply that a library LLM Wiki should be designed as an accountable knowledge environment rather than as a conversational front end. The interface should make pages useful, links meaningful, claims verifiable, uncertainty visible, and correction possible. It should also preserve established library search capabilities while adding new forms of generated organization and exploratory navigation.
The principles also suggest that implementation should be incremental. A library need not begin with a comprehensive generated wiki over all collections. A pilot could focus on a bounded collection, such as an institutional repository, thesis database, local history archive, or course reserve corpus. The important test is whether the system can demonstrate source-addressable claims, contestable organization, and governance workflows within a manageable domain. The next section develops this point further by proposing a research and evaluation agenda for LLM Wiki search.
8 Research and Evaluation Agenda
The LLM Wiki should be evaluated as an information access system, a knowledge organization system, and a governed institutional layer. This requires a broader evaluation agenda than conventional retrieval metrics or chatbot satisfaction ratings alone. A library LLM Wiki may retrieve relevant sources and still fail if it hides provenance, misrepresents disciplinary structure, overstates certainty, reproduces collection bias, or creates unsustainable review burdens. Conversely, a system may generate less polished prose but better support scholarly inquiry if it makes sources, uncertainty, alternatives, and correction pathways visible.
The evaluation agenda proposed here is therefore multidimensional. It combines retrieval evaluation, exploratory search assessment, provenance and citation verification, knowledge organization quality review, bias and omission auditing, trust calibration, and governance analysis. This approach is consistent with research on exploratory search and searching-as-learning evaluation, which argues for measures beyond topical relevance, including learning, engagement, novelty, task outcomes, and user experience (White et al. 2008; Rieh et al. 2014; Palagi et al. 2017). It also aligns with RAG evaluation work that distinguishes retrieval quality from generation quality and emphasizes relevance, accuracy, faithfulness, and grounding (Es et al. 2024).
| Dimension | Guiding question | Possible evidence |
|---|---|---|
| Retrieval effectiveness | Does the system retrieve and surface sources that are relevant, diverse, current, and appropriate to the user's task? | Benchmark queries, relevance judgments, topical diversity, known-item success, source coverage, ranking diagnostics. |
| Exploratory support | Does the wiki layer help users learn a topic, reformulate queries, compare perspectives, and discover useful paths? | User studies, think-aloud protocols, search logs, pre/post topic knowledge, query reformulation patterns, task completion narratives. |
| Provenance and citation verification | Can users and librarians verify which sources support which claims, links, and page summaries? | Claim-source audits, citation accuracy checks, passage-support ratings, provenance completeness, librarian verification tasks. |
| Trust calibration | Do users appropriately trust, question, or reject generated pages based on evidence and uncertainty cues? | Trust surveys, behavioral verification measures, error-detection tasks, confidence calibration, post-task interviews. |
| Knowledge organization quality | Are generated labels, links, topic boundaries, and source clusters accurate, useful, and aligned with professional knowledge organization practices? | Expert review, authority-file comparison, metadata alignment, link validity checks, cluster coherence, terminology assessment. |
| Bias and omission | Which communities, languages, topics, methods, or perspectives are underrepresented, mislabeled, or systematically omitted? | Collection audits, counterfactual searches, subgroup analysis, language and geography coverage, missing-perspective reviews. |
| Governance and maintenance | How much human review is required, who performs it, how corrections propagate, and whether the layer remains sustainable over time? | Correction logs, review time, version histories, unresolved flags, audit outcomes, metadata feedback records, staffing analysis. |
Table 13 organizes the proposed agenda around seven evaluation dimensions. The table also indicates the kinds of evidence that can be collected for each dimension. The key point is that evaluation should follow the conceptual shift advanced in this article. If the LLM Wiki is an intermediary knowledge organization layer, then it should not be evaluated only as a better answer generator. It should be evaluated as a structure that shapes discovery, interpretation, trust, correction, and institutional maintenance.
8.1 Retrieval Effectiveness
Retrieval effectiveness remains necessary. An LLM Wiki cannot compensate for poor retrieval if the underlying sources are irrelevant, narrow, outdated, or unavailable. Evaluation should therefore include traditional and library-specific retrieval measures: known-item success, topical relevance, ranking quality, source diversity, coverage across collections, and sensitivity to user vocabulary. Recent library discovery evaluation work continues to examine relevance, retrieval scope, and discovery-layer quality as core concerns (Ovadia and Lown 2023; Çelik and Çetinkaya 2025). These concerns remain foundational for LLM Wiki search.
However, retrieval evaluation must be adapted to the wiki layer. The relevant object is not only a ranked list. It is also the set of sources used to generate a page, the sources assigned to a cluster, the sources linked to a claim, and the sources omitted from a conceptual pathway. Evaluation should therefore ask whether retrieved materials adequately support the page structure. A topic page generated from an unbalanced source set may appear coherent while misrepresenting the field.
8.2 Exploratory Support and Learning
The LLM Wiki’s strongest claim is that it supports exploratory search. This claim requires empirical testing. User studies should examine whether the system helps users learn unfamiliar topics, identify useful vocabulary, understand subtopics, compare perspectives, and reformulate research questions. Measures might include pre/post topic maps, query reformulation logs, task narratives, source-selection quality, and user reflections on uncertainty and learning.
Exploratory search evaluation is methodologically difficult because success is not reducible to a single correct answer. White et al. (2008) argue that exploratory systems should be assessed through combinations of objective and subjective measures, including novelty, engagement, satisfaction, and task outcomes. For LLM Wikis, this suggests mixed-methods designs: controlled comparisons against discovery search and RAG chatbots, think-aloud protocols, interaction logs, and post-task interviews. The goal is to determine whether users become better oriented, not merely whether they feel assisted.
8.3 Provenance and Citation Verification
Because source-addressability is central to the LLM Wiki, provenance evaluation should be a major component of future studies. Evaluation should ask whether generated claims are actually supported by cited sources, whether the cited evidence is specific enough, whether links between claims and passages are accurate, and whether users can successfully verify claims. RAG evaluation research increasingly emphasizes faithfulness and grounding, but library contexts require more granular verification because scholarly use depends on traceable evidence (Es et al. 2024; Chari et al. 2023).
A claim-source audit could sample generated pages and classify each claim as directly supported, partially supported, unsupported, contradicted, or unverifiable. Citation verification could examine whether references are real, metadata are correct, links resolve, and cited sources support the relevant claim. User studies could then test whether provenance displays improve users’ ability to detect unsupported claims. This moves evaluation beyond whether citations are present toward whether citations function as evidence.
8.4 Trust Calibration
Trust calibration refers to whether users trust the system appropriately. Overtrust is dangerous when generated pages are wrong but polished. Undertrust is also a problem if users ignore useful scaffolding because they cannot interpret provenance or review cues. Evaluation should therefore examine whether users adjust confidence based on evidence quality, review status, uncertainty notes, and source trails.
Trust calibration can be studied through behavioral tasks. Users might be shown pages with supported and unsupported claims, reviewed and unreviewed labels, or strong and weak evidence trails. Researchers can then observe whether users verify claims, revise their judgments, or select better sources. Surveys and interviews can supplement behavioral evidence, but self-reported trust should not be the only measure. The most important question is not whether users like the interface, but whether they use it with appropriate skepticism and confidence.
8.5 Knowledge Organization Quality
The LLM Wiki also requires evaluation as a knowledge organization system. Generated labels, topic boundaries, links, clusters, and entity pages should be assessed for accuracy, coherence, usefulness, and alignment with domain knowledge. This evaluation should involve librarians, subject experts, and where relevant, community stakeholders. Authority files, subject headings, classification schemes, and local metadata can provide comparison points, but they should not be treated as unquestionable ground truth because inherited knowledge organization systems also contain biases and exclusions (Hjørland 2008).
Possible methods include expert review of generated topic pages, comparison with controlled vocabularies, cluster coherence scoring, link validity assessment, and analysis of terminology variants. Evaluation should also examine whether generated structures support user understanding. A technically accurate topic map may still fail if users cannot interpret it or if it does not support their search tasks.
8.6 Bias, Omission, and Representational Harm
Bias and omission evaluation is essential because the LLM Wiki can amplify both model bias and collection bias. If the underlying collection underrepresents certain languages, regions, communities, methods, or epistemologies, generated pages may reproduce that absence as if it were a feature of the field. If a model prefers dominant terminology, it may mislabel local or marginalized knowledge. These risks are especially serious in libraries, where representation is tied to access, memory, and institutional legitimacy.
Evaluation should therefore include bias audits and missing-perspective reviews. Researchers can compare generated pages across languages, regions, disciplines, and communities. They can test whether pages identify gaps in the collection or silently normalize them. They can examine whether subject headings and generated labels conflict in ways that reveal outdated, harmful, or inadequate terminology. The goal is not to eliminate all bias, which is impossible, but to make representational limitations visible and correctable.
8.7 Governance and Maintenance Evaluation
Finally, LLM Wikis must be evaluated for maintainability. A system that requires constant expert correction may be impractical. A system that requires little correction because no one reviews it may be unsafe. Governance evaluation should therefore measure review workload, correction frequency, error recurrence, version history quality, metadata feedback, librarian satisfaction, and policy fit.
Maintenance evaluation should also examine lifecycle questions. What happens when sources are added, removed, corrected, or relicensed? How are stale pages identified? How are high-risk pages prioritized for review? How do user flags move into librarian workflows? How do corrections affect future generation? These questions are rarely captured by one-time usability studies, but they are central if the LLM Wiki is treated as a semi-persistent knowledge organization layer.
8.8 Study Designs for Future Research
Future empirical work could proceed through staged studies. A first study might compare traditional discovery search, RAG chatbot search, and LLM Wiki search for exploratory academic tasks. Participants could complete topic-learning and source-selection tasks while researchers collect interaction logs, think-aloud data, source quality judgments, and post-task interviews. A second study could focus on provenance by testing whether claim-source maps improve users’ ability to verify generated claims. A third study could examine librarian workflows by asking metadata and reference professionals to review, correct, and govern generated pages over time.
The strongest research program would combine system benchmarks, controlled user studies, expert audits, and longitudinal governance analysis. This would allow researchers to ask not only whether LLM Wikis improve immediate user experience, but whether they produce better knowledge structures, more accurate source use, better calibrated trust, and sustainable library stewardship. Such a program would also position the conceptual article and a later empirical prototype study as complementary contributions: the first defines the model, and the second tests its consequences.
9 Discussion
This article has argued that the most important question for generative AI in library search is not whether an LLM can answer a user’s question, but what form of information intermediation the system makes possible. The LLM Wiki reframes AI-mediated search as the construction of an inspectable, navigable, and governable knowledge organization layer. Adjacent work on LLM-compiled wikis for policy retrieval suggests that wiki-like generated structures are emerging as a practical pattern, but the present article theorizes the form specifically as a library and information science problem of intermediation, provenance, and knowledge organization (Zhang 2026). This reframing has implications for academic libraries, information retrieval, knowledge organization, professional roles, and the governance of AI-mediated discovery.
The argument is deliberately positioned between technological enthusiasm and blanket rejection. Generative AI can help users enter unfamiliar domains, summarize source clusters, identify concepts, and move across literatures. At the same time, large language models can produce unsupported claims, reproduce social and collection biases, obscure labor and infrastructure, and create unwarranted trust through fluent language (Bender et al. 2021; Crawford 2021; Ji et al. 2023). The LLM Wiki is not a claim that these risks disappear when outputs are arranged as pages. It is a claim that a wiki-mediated intermediary layer gives libraries a more appropriate design target: not frictionless answer production, but accountable knowledge intermediation.
| Domain | Strategic implication | Research consequence |
|---|---|---|
| Academic libraries | Generative AI search should be integrated with collection stewardship, metadata practice, instruction, and user trust rather than deployed only as a conversational layer. | Study LLM Wikis in bounded library collections and real academic tasks. |
| Information retrieval | Evaluation should move beyond answer quality to include intermediary structures, exploratory pathways, provenance, and maintenance. | Develop benchmarks for claim-source mapping, exploratory navigation, and provenance usability. |
| Knowledge organization | Generated topic pages, links, and clusters should be treated as provisional knowledge organization decisions requiring scrutiny. | Compare generated organization with controlled vocabularies, local metadata, and domain expert judgment. |
| Library professional roles | Librarians become reviewers, maintainers, educators, auditors, and designers of AI-mediated knowledge structures. | Analyze workload, expertise, role boundaries, and professional values in human-AI maintenance workflows. |
| AI governance | AI search systems require policies for versioning, correction, review status, bias audits, licensing, and metadata feedback. | Examine institutional accountability across vendors, libraries, users, and AI infrastructures. |
Table 14 summarizes the broader implications of the argument. The table should be read as a bridge between the conceptual framework and future empirical work. If LLM Wikis are treated as information intermediaries, then their value depends not only on model capability but also on organizational design, professional responsibility, and evaluation methods that can account for source traceability and knowledge maintenance.
9.1 LLM Wikis as Information Infrastructure
The LLM Wiki should also be understood as a proposed information infrastructure. It organizes relations among collections, metadata, sources, users, professional practices, and computational systems. Information infrastructure is not merely technical plumbing; it is a sociotechnical arrangement through which knowledge work becomes possible, durable, and governable (Nowviskie 2020; Brand 2020). From this perspective, the LLM Wiki is not valuable because it adds a new screen to search. It is valuable, if at all, because it reorganizes the infrastructure through which retrieved sources become representations, representations become navigable knowledge objects, and knowledge objects become subject to institutional maintenance.
This infrastructure framing also clarifies the central risk. If an LLM Wiki is treated as a front-end feature, its pages may become influential without becoming accountable. If it is treated as information infrastructure, then provenance, review status, correction workflows, versioning, authority control, and auditability become core design requirements rather than optional quality improvements.
9.2 Implications for Academic Libraries
For academic libraries, the LLM Wiki shifts the AI conversation away from chatbot adoption and toward infrastructure design. Many library discussions of AI focus on reference support, search assistance, information literacy instruction, metadata generation, or operational efficiency (Wheatley and Hervieux 2019; Cox et al. 2019). These are important, but they can encourage a fragmented view of AI. The LLM Wiki instead asks how generative AI might reshape the discovery layer itself: how collections are represented, how users learn from search, how evidence is exposed, and how generated structures are maintained.
This has practical implications for implementation. Libraries should be cautious about adopting AI search systems that provide fluent answers without claim-level provenance, review workflows, or integration with metadata and authority practices. A system that answers questions but cannot show how claims map to sources may be useful for casual orientation, but it is poorly aligned with scholarly search. Conversely, a system that exposes source trails, review status, and correction mechanisms may better match academic library values even if its generated prose is less seamless.
The LLM Wiki also suggests a more realistic pilot strategy. Rather than attempting to generate a comprehensive AI layer over all holdings, libraries could begin with bounded collections: theses and dissertations, institutional repository records, local history collections, open educational resources, or subject-specific bibliographies. Bounded domains make it easier to evaluate source coverage, verify claims, involve subject librarians, and design governance workflows. This incremental strategy matches the article’s claim that LLM Wikis should be built as accountable knowledge infrastructures, not as novelty interfaces.
9.3 Implications for Information Retrieval
For information retrieval, the LLM Wiki expands the object of evaluation. Traditional IR evaluation has often centered on matching queries to documents, while generative IR and RAG evaluation increasingly examine answer relevance, faithfulness, and grounding (Cao et al. 2024; Es et al. 2024). The LLM Wiki adds another evaluative object: the intermediary structure. Researchers should ask whether generated pages, links, clusters, and provenance trails help users search, learn, verify, and revise.
This implies that future IR work should evaluate not only outputs but also paths. A user may not need the best answer immediately. They may need a sequence of useful moves: from vague interest to topic page, from page to source cluster, from source cluster to representative works, from works to methods, and from methods to refined query. The quality of AI-mediated search may therefore depend on navigational trajectories as much as answer accuracy.
The LLM Wiki also invites new benchmarks and tasks. Claim-source mapping, link explanation, page freshness, uncertainty labeling, and source-cluster coherence could become evaluation targets. These tasks would connect IR more directly to information science concerns about representation, provenance, and user learning.
9.4 Implications for Knowledge Organization
For knowledge organization, the LLM Wiki makes generated organization visible as a research problem. LLMs can generate labels, infer relations, summarize topics, and cluster sources. These capabilities may appear to automate parts of knowledge organization, but they also raise familiar concerns in intensified form. Categories can misrepresent communities. Links can privilege dominant perspectives. Generated summaries can naturalize collection gaps. Authority can appear where only statistical association exists.
The contribution of the LLM Wiki is to make these generated structures inspectable and governable. Rather than hiding machine organization inside rankings or embeddings, the wiki layer surfaces it as pages, links, and claims that can be evaluated. This gives knowledge organization researchers a concrete object of study: how machine-generated concepts and relations interact with controlled vocabularies, local metadata, disciplinary traditions, and user understanding.
This also creates an opportunity for libraries to use LLM Wikis diagnostically. Generated pages may reveal metadata gaps, inconsistent terminology, underrepresented topics, or weak authority control. In this sense, the LLM Wiki is not only an interface for users. It can also become a mirror held up to the collection and its descriptive infrastructure.
9.5 Implications for Professional Roles
The LLM Wiki reframes library professionals as stewards of AI-mediated knowledge structures. This role extends beyond prompt writing or tool adoption. It includes evaluating generated pages, correcting claims, auditing bias, maintaining authority mappings, designing user education, documenting policies, and translating user feedback into metadata improvements. This is consistent with emerging accounts of AI-augmented library work that emphasize responsibility for inputs, outputs, evaluation, education, and reflection (Marchionini 2024; Cox et al. 2019).
This shift also raises workload and labor questions. A maintainable LLM Wiki requires human review, but review labor is not free. Libraries will need to decide which pages require expert review, which pages can remain clearly marked as unreviewed, which user reports require escalation, and which errors are tolerable in exploratory contexts. These decisions should not be hidden in vendor defaults. They are professional and institutional choices.
The educational role of librarians may also become more important. Users will need help interpreting review status, provenance displays, uncertainty labels, and distinctions between evidence and inference. Information literacy instruction may therefore expand from evaluating sources to evaluating AI-mediated source structures.
9.6 Risks and Limitations
The LLM Wiki concept has several risks. First, it may create an illusion of stability. A page-based interface can make generated synthesis look more authoritative than a transient chatbot response. Semi-persistence and review status are intended to mitigate this risk, but they do not remove it. Second, the wiki layer may amplify biases in collections, metadata, models, or user feedback. Third, maintenance may become unsustainable if generated pages proliferate faster than libraries can review them. Fourth, licensing and privacy constraints may limit what evidence can be exposed, especially when pages draw on licensed full text or user interaction logs.
There is also a risk of institutional dependency. If an LLM Wiki is built into a vendor discovery system, libraries may have limited control over retrieval logs, model changes, correction workflows, or provenance displays. This would weaken the governance function that makes the concept valuable. Human-centered LLM transparency work emphasizes that transparency must be designed around user needs and actionable oversight rather than treated as a generic disclosure feature (Liao and Vaughan 2024). Libraries should therefore treat transparency, exportability, audit access, and local configuration as procurement issues, not optional features.
Finally, the conceptual argument in this article requires empirical testing. It is plausible that LLM Wikis will improve some exploratory tasks while burdening others. Some users may prefer conversational answers. Some tasks may not require a wiki layer. Some domains may be too contested or underdocumented for safe automated page generation. The value of the framework lies not in assuming universal benefit, but in giving researchers and libraries a richer vocabulary for asking where, when, and how wiki-mediated AI search is appropriate.
9.7 Summary of the Discussion
The LLM Wiki is best understood as a research agenda and design orientation rather than a finished system category. It asks information science to evaluate generative AI search at the level of intermediary structures: pages, claims, links, source trails, review states, and governance loops. This perspective preserves what is valuable about RAG and conversational search while resisting the reduction of library search to answer generation. It also places libraries in an active role: not passive adopters of AI search, but designers and stewards of accountable knowledge intermediation.
10 Conclusion
Generative AI is changing how information systems present search results, but the central challenge for information science is not simply how to produce better answers. It is how to theorize and design AI-mediated systems that preserve traceability, plurality, verification, and accountable stewardship. This article has advanced a theory of AI-mediated knowledge intermediation, with the LLM Wiki as its principal instantiation. Rather than treating retrieval as a hidden pipeline for answer generation, the theory treats retrieval as the basis for constructing inspectable intermediary knowledge organization layers.
10.1 Summary of the Argument
The article began from a distinction between answer-oriented RAG search and wiki-mediated LLM search. RAG systems are valuable because they connect generation to retrieved sources (Lewis et al. 2020; Gao et al. 2023). Yet when the interface centers a generated answer, it can compress complex literatures, obscure retrieval choices, weaken provenance, and leave little behind for review or maintenance. These limits matter in academic libraries because scholarly search is often exploratory, iterative, and evidentiary rather than merely transactional.
The LLM Wiki responds by shifting the object of theory and design. The central object is not the answer, but the intermediary layer: a semi-persistent set of topic pages, source clusters, claim-source maps, related concepts, citation trails, uncertainty notes, and revision mechanisms. This layer is intended to help users move through a field, verify claims, discover alternative paths, and understand how sources support generated structures. It also gives institutions a concrete object for governance: pages and links can be reviewed, corrected, versioned, audited, and connected back to metadata practice.
10.2 Contributions
This article makes three contributions. First, it advances a theory of AI-mediated knowledge intermediation by introducing intermediary knowledge organization as a level of analysis between retrieval operations and user-facing answers. The distinction is not merely technical; it concerns the form of mediation offered to users. RAG generally foregrounds generated responses, while the LLM Wiki foregrounds inspectable knowledge structures.
Second, the article defines the LLM Wiki as a theoretical construct rather than merely a system proposal. Drawing on exploratory search, knowledge organization, algorithmic accountability, and human-AI intermediation, it identifies formal constructs and four core functions: representational scaffolding, exploratory navigation, provenance exposure, and human-AI knowledge maintenance. These functions explain why the LLM Wiki belongs within information science rather than only human-computer interaction or applied AI system design.
Third, the article proposes design principles and an evaluation agenda for responsible implementation. The design principles emphasize source-addressable claims, contestable organization, separation of evidence from interpretation, support for both browsing and precise retrieval, and library governance. The evaluation agenda calls for studies of retrieval effectiveness, exploratory support, provenance verification, trust calibration, knowledge organization quality, bias and omission, and maintenance burden.
10.3 Future Work
The next step is empirical. A follow-on study should implement a bounded LLM Wiki prototype and compare it with traditional discovery search and RAG chatbot search for exploratory academic tasks. Such a study should not ask only whether users prefer the interface. It should ask whether users learn topics more effectively, identify better sources, verify claims more accurately, calibrate trust more appropriately, and understand the structure of a literature more deeply.
Future work should also examine librarian workflows. If LLM Wikis require review, correction, versioning, and metadata feedback, then their feasibility depends on professional labor and institutional policy. Studies should therefore measure not only user benefit but also maintenance cost, review burden, error recurrence, and the kinds of expertise required to govern generated pages. This is particularly important because AI systems can shift labor invisibly unless governance is designed as part of the system from the beginning (Cox et al. 2019; Marchionini 2024).
Finally, future research should investigate domains where LLM Wikis may be inappropriate or risky. Topics with sparse collections, contested terminology, sensitive community knowledge, licensing restrictions, or high risk of representational harm may require stricter review, narrower generation, or no generated pages at all. The framework advanced here should therefore be treated as a guide for careful inquiry, not as a universal prescription.
10.4 Closing Statement
The value of the LLM Wiki concept is that it changes the question. Instead of asking how libraries can attach an LLM to search, it asks how libraries can govern AI-mediated knowledge intermediation. Instead of asking whether a system can generate an answer, it asks whether the system can help users inspect, navigate, verify, and maintain the knowledge structures through which scholarly search occurs. This is the central shift: from answers to intermediaries, and from generative fluency to accountable knowledge organization.