Typically, a literature review starts with reading many papers and gaining an understanding of the subject matter. However, AI enables a new way of starting your literature review. That is, by understanding concepts, their connections, and mentions within papers. In this tutorial, I would like to show you how you can use AI to convert a set of PDFs into a Wikipedia-like structure that you can navigate and learn from. This method does not replace understanding, but can help you get your foot in the door and try to learn new topics much more quickly.
Learn which tools can speed up your research: This free 21-day email series introduces you to a range of academic tools and how they work together to speed up the time needed for a literature review. Most of these tools have been around for less than a few years, some use AI. Knowing them gives you an almost unfair productivity advantage for your research.
What is a Concept-Network?
A concept network consists of three things: concepts, claims, and links. Concepts at the top are a clean, deduplicated list of the ideas your literature addresses. Each concept consists of claims linked to individual papers (I call them atomic sentences). Finally, concepts and claims can be linked together. (e.g. the statement “Penguins live in Antarctica” links these two concepts together). Here is an example:
How does it help you to learn faster?
From Nesbit and Adesope’s (2006) meta-analysis of 55 studies, we know that using concept maps (or mind maps) produced reliable retention gains across grade levels and disciplines. The largest gains are obtained when learners build the maps themselves, but importantly, even simply studying pre‑constructed concept/knowledge maps still improved retention. If maps can be generated in seconds using AI, we can leverage this, and by manually editing these maps, we can find a trade-off between speed and learning quality.
The reason why this works can be found in Craik and Lockhart (1972) three-tiered model of memory. This model proposed that new ideas are memorised at three levels of depth: structural, phonemic and semantic. Each layer improves retention.
They asked participants to memorise concepts by posing questions designed to activate each tier or depth. The three layers are:
- Structural: Read the word ELEPHANT, highlight it, ask whether it is in capitals, and you might still forget it in minutes. That is why highlighting Papers in Zotero is the wrong way to read papers (see how to properly read PDFs).
- Phonemic: ELEPHANT rhymes with extravagant. Recall will be better, which is also why ancient works (like the Vedas) use rhyme and cadence.
- Semantic: Ask whether elephants are bigger than horses, and what their most distinguishing feature is. Using these questions during memorisation increased recall by 2-6 times.
This means that to anchor concepts, we must see them in relation to other concepts. Mind mapping forces you to do just this, thereby embedding learning in this semantic bottom layer. You cannot draw an arrow from elephant to horse without first deciding what this means and how they connect. While learning happens when you create these maps, you can still use AI-generated ones as a first step to learning, from which you start editing and building on these maps. AI provides the scaffold, but you fill in the details.
Building a Concept Network With AI
To build our concept network, we will use Obsidian, a tool that functions like a personal Wikipedia for exploring connected text files. It is the perfect substrate for building personal knowledge databases because it is based on plain text, is easily accessible to AI, natively processes PDF documents, and allows adding and exploring links. Check out the absolute beginner tutorial on Obsidian, if this is new to you.
Step 1: Collect The Papers
To get started, you will need a few papers that contain the main concepts of the subject you want to write a literature review on. Ideally, you should select frequently cited review papers published in rigorous journals. You can use Consensus as your starting point, or Litmaps/ResearchRabbit to identify impactful papers. Experiment with the number of papers, but since review papers are often quite extensive, selecting just 4-5 might be enough to start.
In the example, I will be using just two PDFs.
Step 2: Install and download the ea-obsidian-world-builder skill
This skill extracts claims and concepts from papers. Each concept will be written into a separate file, and the claims and findings of each paper will be mentioned in the source note for each paper. This follows the note-taking system I have developed over 4 years of academic research.
Step 3: Run the skill
You don’t need to specify much to run the skill. Just upload your papers to Claude (or Codex) and mention the skill. I suggest using the newest and biggest module when dealing with anything involving the synthesis of large bodies of work. In this example, I used Anthopic’S “Opus 4.7”.
Running the skill will take a few minutes and consume a good deal of tokens, especially when creating the knowledge network from scratch. During the building process, the skill will ask you for feedback and opinion. Most importantly, you need to filter out concepts that seem irrelevant to you or redundant to speed up the process.
Step 4: Filtering Concepts
By default, the AI will not really know where your evaluation is headed. It might sometimes identify general concepts such as meta-analysis or effect size, which are things you might already be familiar with. Carefully read through the suggested list of concepts and tell the AI which to eliminate and which to combine. This will make the result much more meaningful.
You can, of course, introduce concepts of your own.
Step 5: Open the resulting folder with Obsidian
Download Obsidian, and open the resulting folder with Obsidian to start exploring it. The result will be a tightly connected “world” of concepts mentioned in these two papers. You can start exploring it and learning from it.
If you are interested, please download the Obsidian vault I generated using the two papers on knowledge network learning mentioned above.
Step 7: Read and understand the network
Remember the finding from Nesbit and Adesope’s (2006) meta-analysis that the greatest gains in learning from mind maps were achieved by those who created them themselves. Just transforming a few PDFs into a knowledge network does not guarantee learning, but it helps you. It provides you with a tool that can greatly enhance it. Think of it as a bicycle. When using your bike, you’re faster than walking, but different rules apply. You need different gear, and it’s not usable in every weather. Similarly, here, if you’re trying to deeply learn a topic, create this knowledge network from scratch by hand and read these papers.
Using GitHub to avoid content loss
If you’re working side by side with an AI agent that edits your notes, it can quite easily happen that things you have created get overwritten, especially when the agent has to make substantial changes. A very simple solution is to put your vault into a local Git repository. You don’t need to upload it to a cloud or pay for git to use this. Here is how:
- Get a free client like Sourcetree
- Drag and drop your folder into Sourcetree to create a repository. All this does is create an invisible “.git” folder containing some config files.
- Make a first “commit”, by selecting all files and clicking the commit button at the top left. This simply marks the state of these files and starts tracking any future changes.
After making some edits and returning to your Sourcetree app, you will see the changed highlights and individual edits file-by-file:
This is especially useful if you use further AI agents to edit your knowledge network.
Using the Knowledge Network to retrieve information
I suggest installing Gemini Scribe as an Obsidian plugin, as well as Smart Connections. These two AI tools will allow you to use AI to retrieve information and process knowledge within it. They are useful if your goal is not learning the subject but more creating a repository of knowledge that you can easily query for specific topics to get evidence-based answers.
Alternatively, you can set up Claude to work with your vault.
Summary & Download Links
This tutorial shows how AI can help you start a literature review by transforming a few PDFs into a navigable “concept network” inside Obsidian. Instead of only reading papers linearly, you explore concepts, claims, and links between ideas in a Wikipedia-like structure that helps you learn new topics faster. Concept maps are a scientifically proven way to learn faster. The single steps are:
- Convert papers into interconnected concepts and claims
- Build a personal “knowledge world” for a research topic
- Use AI as a scaffold for faster semantic learning
- Explore and query the network using AI plugins
- Protect your notes from AI mistakes using Git version control
The goal is not to replace understanding, but to accelerate it. AI generates the initial structure, while real learning happens as you edit, connect, and expand the network yourself. Personally, I use this method to understand new topics at an academic level, without spending too much time and outsourcing too much interpretation to AI.
Download the demo vault I built for this article here.
