More efficient, more accurate, more transparent: Perplexity AI, a tool in the service of scientific research.

Olvasási idő: 6 perc

In recent years, the development of artificial intelligence (AI) has significantly affected web search services. AI-based search engines do not simply return links but summarize and contextualize the results, assign specific sources to each statement, and provide intuitive communication with the user. New-generation search services, such as Perplexity, can deliver relevant results by considering user intent and context, significantly improving the search experience. In this blog post, we present how these advanced search engines work, the technologies behind them, and the advantages they offer in everyday life and academic research.

Introduction

Today, one of the most significant areas of technological advancement is artificial intelligence (AI), which aims to create systems capable of mimicking human thinking, learning, decision-making, and problem-solving. This must be realized in such a way that users can communicate with these services using natural language, often formulated imprecisely. However, the expected results must be produced quickly, and the answers must be accurate. The quality of these responses largely depends on the quality of the data used to train the underlying language model. AI-based services often attempt to provide answers even when they have no relevant information for a particular question.

After the emergence of AI-based chatbots, it became apparent that the next step would be to reference their answers and ensure access to up-to-date information. A key to this was allowing AI assistants to “step out” onto the web and search for real-time information.

Large Language Models (LLMs) and hallucination

Initially, users of large language models did not have access to answers generated from continuously updated web content. Due to the limited data updates of the corpus used by language models and their “need to answer,” even the most popular services are prone to hallucination—generating answers based on previously learned patterns, even to the most absurd questions. This situation appears to be improving, as reasoning models refine statistical response generation with multi-step derivations, logical steps, self-checking, and a focus on coherence. Access to fresh web content also reduces the likelihood of misinformation, although evaluating the reliability of sources remains a challenge.

Perplexity is a search service that not only returns links but also summarizes and contextualizes results and assigns specific references to its claims.

Features

Search interface

The interface itself is not particularly visually stimulating—without a Pro subscription, the service offers an automatic search mode. We can upload files into the search bar and extend or limit the search range to the uploaded file, the web, scientific publications, or social platforms. Additionally, the service recommends a few current news stories and summarizes them in an automated manner.

The search capabilities truly deepen with a Pro subscription, where we can choose from multiple language models: a modified, uncensored DeepSeek R1; Grok-2; Gemini 2.5 Pro; Claude 3.7; GPT-4.1; and o3 mini reasoning models.

A recently introduced option is “Deep Research,” which aims to generate detailed answers to complex questions. The system analyzes the questions to identify user intent and necessary information, then processes numerous sources in parallel. It then organizes the results hierarchically and creates structured reports that can be exported or shared with other users. This is a far more time-consuming process than a typical web search, but the system notifies users with pop-up alerts when tasks are completed.

After answering questions, Perplexity also suggests related questions, which can be triggered with a single click. Answers can always be shared and exported in PDF format, with source references included in the exported material.

For each search thread, we can search for related images or videos or generate images based on preset or custom definitions. There is also an option to rewrite a response if it does not meet our expectations.

Library

The Perplexity Library feature first displays the searches or “topics” we have completed so far, so we can resume our research at any time by reopening a previous topic. However, this menu hides far more functions after clicking the “+” button:

Pages: allows us to create and share articles, reports, or guides on any topic with just a few clicks. The structure, depth, and language of the content are easily customizable and can be expanded at any time with additional paragraphs or existing media. With a corporate subscription, we can build an institutional-level knowledge base using the Internal Knowledge Search feature. This helps combine internal information with web sources to make decisions based on the current trends.
Spaces: One of the most interesting features of Perplexity. We can group our searches, upload files as information sources, and define URLs that we want to use as mandatory sources. We can initiate collaboration and share our content with colleagues.
The system can also be set to answer questions solely based on files uploaded to Spaces or specific websites, thereby creating a personalized or project-level assistant.
For example, we can create an online assistant that only answers questions based on the information sources we specify. After uploading numerical data (Excel or CSV files), the assistant can even generate statistical summaries.

Discovery

It is primarily a news recommendation function that categorizes news into different topics and highlights the most important developments. This feature summarizes the same news from different portals to provide as balanced information as possible on current events.

Prompt or search query?

Perplexity operates as a hybrid between search engines and chatbots. It can handle both detailed, instruction-like prompts and traditional keyword-based searches. The choice depends on user intent; prompts are recommended for complex, multi-faceted answers.

Which LLM model should I choose for searching?

In short, the model choice depends mainly on the type and goal of the question. Reasoning models are best for step-by-step analysis (e.g., math or logic problems) because they can derive solutions and provide detailed explanations. For academic research—like writing papers or complex projects, the Deep Research mode is recommended. This mode can analyze hundreds of sources and produce detailed and referenced reports.

For everyday, quick questions, Automatic or Pro mode is sufficient—the latter provides access to the newest large language models (e.g., GPT-4.1, Claude 3.7) and provides more detailed responses. To find the best fit, it is worth experimenting with multiple modes; experience will be the best guide.

Scientific research and perplexity

Perplexity AI’s Academic Research feature offers a special Focus mode that limits searches exclusively to scientific sources. It uses journals, conference materials, preprint databases (e.g., arXiv), and other open research databases to do so.

For paywalled articles, the system usually processes only metadata and abstracts but indicates where the original source can be accessed. This mode is ideal for launching research, reviewing sources, and organizing them, as every claim is tied to a specific reference, making it easy to verify answers. It is worth combining this with the Deep Research mode; however, as with all AI-based tools, verifying the credibility of the information remains essential.

Which key steps of academic research does the system support?

Perplexity primarily supports literature review preparation; it can identify relevant sources and summarize key trends on a given topic. Within minutes, it can generate a report analyzing hundreds of sources. Narrowing and refining research questions is also efficiently achievable, especially if the search is not initially limited to scientific sources.

The system compares data from different sources, detects patterns, and helps resolve contradictions. This allows us to create comprehensive reports that would take hours or days to produce manually using conventional methods.

Search threads can be shared, thereby accelerating group work. Finally, the Spaces function allows us to create custom assistants that respond based on our own materials and documents—including full-text publications.

Searching for literature or synthesizing sources?

Let’s suppose we are looking for an answer to the following question:

“I’m looking for peer-reviewed English-language literature from the past 3 years about the application of AI in academic libraries.”

After turning on the “Scientific” source filter, Perplexity provides a detailed, structured overview as follows:

Introduction – The impact and challenges of AI technologies (e.g., NLP, machine learning) in libraries.
Technological Background – AI development in library contexts.
Practical Applications – Information retrieval, automation, personalization, education.
Challenges – Ethics, access, training needs.

Additionally, the system provides a list of 60 sources with full-text access and can even include video sources. The search can then be refined and deepened by further questions.

Data protection – what you should know

By default, the provider uses searches for model development, but this feature can be disabled in settings—even with a free account. After account deletion, the data are erased within 30 days or sooner upon request.

Infallible?

In general, Perplexity is not infallible. Similar to most AI-based services, Perplexity may provide irrelevant or incorrect answers, especially for less precisely formulated search queries. Therefore, it is important to verify the sources of the data provided and validate their authenticity. The use of AI also raises ethical questions: Who bears responsibility when AI makes a mistake? How can we trust a system whose functioning—due to its black-box nature—we do not fully understand? Given all this, responsible use and adherence to existing institutional regulations are especially important.