A general assistant
Answers from training data and whatever it can fetch. Sources are recalled or searched for at the moment of answering. You verify afterwards, if you can work out what to verify.
AI-assisted research
AI for Academic Research
A general chatbot answers from whatever it absorbed during training. Nodus asks a model to work on something much narrower and far more useful, namely the documents, sources and notes you brought in. Every interpretation it produces is stored with the passage behind it, so you can check the output against the page instead of taking it on faith.
01 · The difference
Grounded in your corpus, not in the internet.
The failure researchers know best is a fluent answer with a citation that turns out not to exist. It happens because a general assistant is asked to produce a scholarly sounding paragraph without having any particular scholarship in front of it.
Nodus turns that around. You assemble the corpus first, from your Zotero library, from bibliographic exports, or from your own files, and only then is the model asked to work inside it. When Nodus records a claim, it records which work it came from and which excerpt supports it, along with the page where the source provides one. A citation cannot be invented, because the thing being cited is an item already sitting in your Library.
That is the whole idea. The AI proposes, the evidence stays attached, and you decide. Nothing in Nodus asks you to trust a paragraph on its own merits.
Nodus
Answers from a corpus you chose and imported. Each stored interpretation carries its work, its excerpt and its page. Verification is a click, because the preserved original is already on your disk.
What neither can do
Judge whether an argument is sound, or whether a literature is worth building on. Both of those are your job, and Nodus is built to make that job easier rather than to take it over.
02 · What the AI actually does
Specific jobs, not a single magic button.
AI in Nodus is split into separate tasks. Each one has a narrow brief, its own output shape and, if you want, its own model.
01
Extracting ideas from a work
Analysis reads a source and proposes units of five kinds, claims, findings, constructs, methods and frameworks. Each one is recorded with the excerpt that supports it, and with a note of whether that excerpt is quoted or paraphrased. If only an abstract was available, you get correspondingly little back, and Nodus says so rather than padding the result.
02
Relating and contrasting them
Ideas are connected with relations that carry a meaning, such as supports, contradicts, refines, extends or applies to. A relation can also record a shared method or a precondition. Each connection notes whether the source stated the relation or the analysis inferred it, so an inference never passes itself off as a quotation.
03
Surfacing debates and research gaps
Disagreements between sources become debate views grouped by stance. Unanswered questions, acknowledged limitations, flagged future work and unresolved contradictions become gaps, each with the evidence that exists and a note of what is still missing. Coverage analysis then measures how well the corpus supports a theme, which is what tells you whether a gap is real or simply an area you have not read yet.
04
Semantic search
Embeddings let you find the passage that means what you asked, even when it uses none of your words and is written in another language. This is the one AI task that runs across everything, and it can be computed entirely on your own machine.
05
Long-form synthesis
Deep Research runs as a queued job over the scoped corpus and returns a structured report full of citations that open the source they point at. Immersion builds a guided route through a research question. Both are drafts written for a reader who will check them, and Nodus presents them that way.
06
Hypotheses and writing
A hypothesis pairs a statement with the ideas that support it, the risks against it and a proposed test. Those are separate fields on purpose, so speculation cannot pass for a result. In the writing workspace, drafts are composed beside the corpus with citations that resolve to real items in your Library.
07
Reading what a scan hides
Vision-capable models can read pages as images, which is how scanned documents, figures, tables and formulas get in when a plain text layer misses them. The OCR Workspace puts that behind a page by page review, so nothing enters your corpus before you have looked at it.
03 · Choosing models
You pick the provider, and you can pick it per task.
Nodus has no model of its own and no bundled subscription. It talks to the providers you configure with your own credentials, and none of them is required for the application to run.
For text, Nodus supports Anthropic, OpenAI, ChatGPT · Codex, GitHub Copilot, OpenCode Go, OpenRouter, Groq, Cerebras, DeepSeek, Google Gemini, Xiaomi MiMo, and the local runtimes Ollama and LM Studio. For embeddings it supports OpenAI, Google Gemini, OpenRouter, Ollama, LM Studio and a small multilingual model that Nodus manages and runs itself.
Model settings come in two modes. In the simplified mode you make one choice and everything follows it. In the advanced mode each task gets its own selector, because they want different things. Extraction runs many times over long documents, so a cheap and reliable model pays off. Long-form synthesis rewards a strong model. The small deduplication and relation calls during a deep scan reward a fast one. Vision, chat, Deep Research, Immersion, writing, the argument map, author synthesis and hypotheses each keep their own choice, so changing one never quietly retargets another.
Nothing runs until you say so. Model-assisted features stay inactive until a provider or a local model has been configured, and the application is perfectly usable without one. A corpus, its documents, exact search, notes and writing all work with no AI at all.
04 · Local models
Research material that never leaves the machine.
Unpublished work, confidential interviews and material you are not licensed to send anywhere are ordinary facts of academic life. Nodus can run its AI features completely offline.
Ollama and LM Studio
Point Nodus at a local runtime and the same tasks run against a model on your own hardware, whether that is extraction, synthesis, chat or embeddings. No key, no request leaving the machine, no provider terms to read.
Embeddings on device
Semantic search does not need an external provider at all. Nodus can manage a small multilingual embedding model itself, and the resulting index lives on your disk with the rest of the vault.
The honest trade-off
Local analysis is slower, often a great deal slower, and smaller models are less reliable at returning the structured output the analysis pipeline expects. That is a real cost rather than a footnote, so it is worth trying a model on a handful of sources before you commit a whole corpus to it.
Mixing is allowed
Because the tasks are separate, a common arrangement is local embeddings and local extraction, with a stronger hosted model kept for long-form synthesis. Some people do the reverse, depending on what their material allows.
05 · Limits
What this does not promise.
A research tool that oversells its AI is worse than one with no AI at all, because it teaches you to stop checking.
Nodus does not automate research and does not replace the researcher. It cannot guarantee that an extracted idea faithfully represents its source, that a proposed relation holds, or that a generated report reaches a correct conclusion. Extraction is a first pass over your material, and it can be wrong.
What Nodus does instead is make checking cheap. Every idea carries its excerpt. Every excerpt points at a preserved original you already have on disk. Report citations open the work they refer to, and an audit ledger records what happened in the vault. The evidence audit in the Zotero plugin works the same way, flagging claims whose support looks thin and sending you back to the sources.
Where AI deliberately does not go
The teaching vault generates teaching material only. Nodus does not send student rosters, grades or answers to a model, and does not use AI to grade, profile or evaluate students. Nodus Protect processes documents entirely on your computer and never sends them to a provider. In the databases vault, AI columns work with the fields and values actually present in your rows.
A note on academic use
Whether AI assistance is acceptable in a given piece of work is for you, your supervisor, your journal and your institution to decide, not for a piece of software. What Nodus does is keep the provenance you would need in order to answer that question honestly.
Keep going