AI-assisted Literature Search

How AI-supported scholarly search services work

As AI technologies have rapidly developed in recent years, new types of search services for scholarly information have emerged, services such as Scopus AI, Elicit, Undermind, Keenious, SciSpace, Google Scholar Labs, Web of Science Research Assistant, Consensus, and Asta. Unlike traditional databases and search services, these tools allow users to ask questions in natural language. The result is not only a list of publications, but also a summary explaining how the retrieved articles answer the search question.

Although these services differ in design and functionality, they are generally built on similar underlying technologies combined in different ways. These technologies are not unique to scholarly search services; the same methods can be used in other situations where information needs to be retrieved and summarized. In scholarly search services, however, they are applied specifically to collections of academic information. The following sections provide an overview of some of the core technologies involved. Large language models (LLMs) often play a central role in several of these steps, commonly using models from providers such as OpenAI, Google, or Anthropic.

The search query

One obvious difference compared with traditional information retrieval is that the question is formulated in natural language rather than as a search string constructed from keywords. For example, a user might write: What factors influence students’ motivation in distance learning?

Some services use a language model at the very beginning to analyze the question and suggest clarifications or elaborations, while others perform the search directly. The search process itself may vary, but a common approach is to combine a classic keyword search with semantic search. This combination is sometimes referred to as a hybrid search.

In the keyword-search component, a language model translates the user’s question into a conventional search string. The example above might become something like:

((factors OR determinants OR variables OR elements) AND (students OR learners) AND (motivation OR engagement OR drive OR incentive) AND (distance learning OR online learning OR remote education OR e-learning))

This search string is then run against a search index and ranked using methods similar to those used in traditional search services.

Semantic search, by contrast, attempts to understand the meaning of the question and match it with articles that discuss similar concepts, regardless of the exact wording used. One common method used in AI-supported search services is vector search. The query and indexed articles are transformed into numerical representations known as vectors, which encode semantic meaning. Vectors are considered to represent similar concepts if their values are close to one another, allowing the system to identify documents that are conceptually related to the search question. These vectors are typically generated from titles, abstracts, and subject headings, although full texts may also be used.

Both search methods are used because they complement one another.

• Keyword searching is computationally efficient and works well for precise terms and names, but it may miss relevant articles that use different terminology.

• Semantic search can identify relevant articles even when the wording differs substantially, but it is computationally more demanding and may sometimes miss exact matches or interpret the query too broadly.

Ranking

The results from the different search methods are combined by assigning scores to articles based on their ranking within each result set. These scores are then merged into a single overall ranking.

Some services also perform an additional reranking step. In this process, the search query and the highest-ranked articles are sent together to an AI model, often an LLM, which evaluates the relevance of the results in greater depth. This can improve ranking quality but is computationally expensive and is therefore typically applied only to a smaller subset of the retrieved documents, such as the top 100 results.

Citation chains

Some services also use citation chaining, sometimes referred to as citation snowballing. In this approach, the system examines the references and citations connected to already retrieved articles to identify additional relevant publications.

It is worth noting that there are some services primarily focusing on citation relationships rather than using AI-based retrieval. Examples include Litmaps, Research Rabbit, and Connected Papers.

Search results

A characteristic feature of AI-supported search services is that results are presented both as a publication list and as a generated summary explaining what the retrieved literature says about the user’s question.

These summaries can range from short overviews to extensive literature reviews and reports and are based on a method known as Retrieval Augmented Generation (RAG).

In a RAG workflow, the retrieved articles, often titles and abstracts, are sent to a large language model together with the original search question and a prompt stating that the model should base its responses solely on the attached articles. The language model then generates a synthesis of the articles in relation to the question, often with references indicating which articles support specific claims.

Because the generated response is based on retrieved documents, users can inspect and verify the sources themselves. The method also reduces the risk of fabricated references, a problem that can occur in general-purpose AI chatbots that generate responses directly from training data without retrieval.

Further searching

The combination of retrieval and language-model technologies also enables AI-assisted search services to support iterative searching. In a traditional database, users manually reformulate and refine their searches as they evaluate the results. AI-supported services, by contrast, often suggest follow-up questions or propose additional searches based on the retrieved literature.

These suggestions are generated by language models that have access to both the original query and retrieved documents. Based on this information, the system can identify aspects, perspectives, or gaps that may be relevant for further exploration.

Deep research

Some services go beyond a single search-and-summary cycle and instead devote more computational resources to refining the search process and producing more comprehensive outputs. This is sometimes referred to as deep research and typically relies on agent-based AI systems, where language models operate more autonomously through multiple iterative steps.

For example, the system may conduct an initial search, analyze the results, identify missing aspects or gaps, formulate new queries, and perform additional searches. Citation chaining and reranking are often incorporated into these workflows.

The synthesis stage is also more ambitious. Retrieved articles may be used to generate long-form reports, literature reviews, evidence tables, timelines, visualizations, and analyses of research gaps. Some systems validate citations against underlying databases and may even execute code to create tables and visual outputs. However, the quality of the final report still depends fundamentally on the quality of the retrieval process itself: if important articles are not found, the resulting synthesis will also be incomplete.

These advanced methods are resource-intensive and may take several minutes to complete, but they can produce significantly better results in terms of relevance and coverage.

Data sources

Regardless of the retrieval method used, no service can retrieve publications that are absent from its underlying data sources. The available data varies between services.

Services such as Scopus AI and Web of Science Research Assistant search for their own proprietary databases, while Google Scholar Labs builds data indexed by Google Scholar.

Many other AI-supported scholarly search services, including Elicit, Undermind, Keenious, and Asta, primarily build their indexes from Semantic Scholar or OpenAlex, two large open collections of scholarly metadata. These in turn aggregate information from sources such as Crossref and PubMed.

Most services primarily search for metadata — titles, abstracts, and sometimes subject headings — rather than full texts. As a result, generated summaries are usually based on abstracts rather than complete articles. Some services do index full texts, but often only on a limited scale. In any case, this means that results differ not only because of variations in retrieval methods, but also because the services search for different collections of material.

AI-supported information retrieval versus traditional search services and AI chatbots

AI-supported scholarly search services, traditional databases, and general AI chatbots each have different strengths and weaknesses when it comes to information retrieval. As a user, you need to consider how to use them based on your own needs.

The strengths and limitations of the different types are described below. It should also be noted that the boundaries between these types of services are somewhat fluid: established databases are incorporating more AI features, general chatbots are becoming increasingly “search-like,” and AI-supported scholarly search services are partly based on the same technologies as traditional search systems.

AI-supported search services for information retrieval

One advantage of AI-supported scholarly search services is that they often have a low entry barrier: you do not need to construct a perfect search string or know the correct subject headings but can instead enter the search query directly in natural language. The services use technologies to understand the meaning behind the question, which allows them to retrieve relevant articles even when the articles use different terminology than in your query. Terms you then discover in the results can also be reused to build better searches in traditional search systems.

In addition to the results list itself, most AI search services generate a summarised answer based on the search query, with references to the retrieved literature. This can provide a quick overview of what the research says and is particularly valuable in an early exploratory phase. The services often include opportunities for dialogue to ask follow-up questions and receive suggestions for new angles, which is also useful when the research question is not yet fully defined.

Some services such as Undermind, Elicit, and Asta (Ai2) have agent-like, iterative workflows and take this a step further by working more autonomously with the search query: they can perform several searches in sequence, follow citations, and adjust search terms based on what they find. In this way, they can identify key articles on specific research questions, even across very large databases.

One of the engines behind these functions is usually a large language model such as GPT, Claude, or Gemini. The difference compared to using these providers’ chatbots directly (see below) is partly that the scholarly AI services search within a predefined scholarly information set and that they use a specialised, preconfigured workflow for how the language model is applied and what instructions it receives.

The strengths have a downside: AI-based services are not fully “deterministic” in their processing of search queries and ranking, meaning that the same query may produce different results at different times. This is a particular limitation in contexts requiring reproducibility, such as systematic reviews. It is rarely possible to understand in detail why a given document has been ranked as relevant. The services also only display a subset of results, and the selection process is not fully transparent, meaning you cannot control what is excluded.

Data sources differ between services and are not always clearly documented, but many services are based on the open indexes Semantic Scholar or OpenAlex. The use of these data sources leads to a certain standardisation and results in weaker coverage of, for example, books and humanities content. It is also common that services only have access to article abstracts, not full texts.

Even when the underlying sources are high quality, it is still possible that the generated summaries are not fully accurate: they may misinterpret or oversimplify what the articles actually say, and the well-formulated answer with citations may easily appear more authoritative than it is. You therefore need to critically review and follow up.

A general comment on these services is that they are developing and changing rapidly. New features are added, underlying models are replaced, and indexes are expanded, which of course also affects the predictability of their performance.

Traditional search services

In traditional search services and databases, you define exactly what is searched and how. You can use controlled subject headings, build Boolean search strings with AND/OR, and apply filters such as author, journal, or publication year, and you know afterwards which terms actually matched. The same search string also produces the same results every time it is run (assuming the database has not been updated), which is a requirement for systematic reviews and other contexts where reproducibility is needed. Traditional search services also display all results and allow you to browse through them. When you need to ensure that no relevant studies are missed, they remain the most reliable option.

However, this level of control can also mean a higher threshold for achieving good and comprehensive results. The search query must be translated into a functional search string, which may require Boolean logic and familiarity with relevant subject headings and terminology. The search engine primarily matches character strings and does not directly understand the meaning of your query. Although there may be some support for synonym expansion, you often need to identify, think of, and test multiple synonyms and variations yourself. You also do not get an automatic summary or synthesis for overview purposes but instead need to open and read abstracts from a large number of articles.

For certain simpler search tasks, traditional search services are the obvious choice, for example when quickly retrieving specific titles for a reference list or checking how to access a particular publication.

Traditional search services often also provide richer filtering options and make it easier to refine searches according to your needs.

The traditional services subscribed to by libraries are often more readily accessible via library systems through link resolver and proxy integrations, although similar solutions may also exist in newer AI-based services.

AI-supported scholarly search services are often limited to certain broad, article-focused datasets. Equivalent broad search services also exist among traditional systems (such as Scopus, Web of Science, and Google Scholar), but in addition there is a wide range of other content: specialised databases, statistical sources, book and library catalogues, national sources, digitised material, and more. In these systems, it is usually also clear what data is included, how the material is indexed, and what selection criteria apply.

General AI chatbots

General AI chatbots (ChatGPT, Claude, Gemini, Copilot) are general-purpose language models trained on vast amounts of text from the internet. Their primary purpose is to generate and process conversational text and answers, not to function as search engines for scholarly information.

In an information-seeking context, chatbots can serve as versatile assistants for explanations, brainstorming, or identifying search terms for research areas, but they have limitations when it comes to actual information retrieval.

A chatbot has no academic index. If it searches, it searches the open web, and the response may be based on everything from news articles and blog posts to Wikipedia and occasional scholarly articles. It is not guaranteed that the answer is grounded in scholarly literature. Unlike an AI-supported scholarly search service, which always searches within its academic database, the chatbot itself decides whether to search at all or whether it can answer directly from its training data, which also has a fixed cutoff date and therefore lacks the most recent research.

When the model attempts to “guess” an answer without proper grounding, it may also fabricate references to studies that appear real but do not exist. Although this has improved, the risk is still significantly higher than in dedicated AI search tools connected to scholarly databases. The references produced are therefore less reliable and require greater critical scrutiny than those returned from a dedicated scholarly service. Chatbots also lack research-specific features such as filtering options, reference export, and analysis of citation networks.

At the same time, the pace of development is high, and several services now include dedicated research modes that perform more systematic web searches and compile supporting material. In addition, MCP standards are beginning to make it possible to connect general AI chatbots to scholarly data sources. The boundary between chatbots and dedicated scholarly search tools may therefore become less distinct over time.