NodusBeta
Self-hosted research-paper analysis
The problem it removes
Existing tools answer the easy question. You get a readable summary with citations attached, and it holds up until someone asks where one particular sentence came from. Nodus treats provenance, conflict and trust as structured output rather than prose: a claim belongs to named papers, a disagreement has a typed cause, and a quality tier has a derivation you can inspect. The report is still readable. It is now also checkable.
How a run works
Three stages, and minutes rather than seconds, because it is fetching and reading real papers. Retrieval pulls around twenty ranked papers from Semantic Scholar. Extraction takes up to twelve claims from each, embeds them, and groups equivalent claims across papers while keeping every source claim, its stance and its extraction confidence. Synthesis assembles sections as they finish, each citing the clusters underneath it. Progress streams phase by phase, and follow-up questions are scoped to a previous query and linked to their parent, so the refinement chain stays inspectable.
Judgement stays with the researcher
The formula does not know that a trial's registry entry was amended, or that four corroborating cohorts draw on the same registry. You do. Override the tier, give the reason, and both the computed value and your correction stay on the record. Edited clusters and reports are pinned, which means re-analysing a question builds on your review rather than discarding it.
Runs on your hardware, with your model
Nodus is self-hosted. Papers, claims, clusters and reports live in your own Postgres with pgvector; there is no hosted service holding them and no accounts to create. The model is swappable between Gemini, Anthropic and Ollama, and on Ollama the entire pipeline runs locally with nothing leaving the machine, which is the configuration for unpublished or embargoed work. Setup is a clone, one env file, a migration and a uvicorn process.
Open source, including the arithmetic
The ranking weights, the quality formula and the clustering logic are all in the repository, which is the point of shipping it this way. Read the code, disagree with a weight, and change it for your own field. Releases are published on the repo.
What's in the box.
Everything included out of the box to run this solution reliably in production.
Lineage for every claim
Each cluster carries the chronological chain of papers behind it: the originating paper and its year, the span of years the claim survived, and how each later paper relates to it. Relationships are typed as origin, supports, contradicts or extends, so you can see whether the evidence accumulated or split apart.
Typed disagreement drivers
Spotting a conflict is easy; Nodus records what is driving it. Every cluster gets drivers classified across eight categories, from methodology and population to metric definition and publication bias, and each driver names the specific papers on either side.
Quality tiers that are arithmetic
A tier comes from a published formula: study design 40 percent, sample size 20, corroboration 20, extraction confidence 20, less a conflict penalty of up to 15. No language model rates the evidence, and every input is shown so you can redo the sum yourself.
Retrieval with the weights written down
Around twenty papers per question come from Semantic Scholar, ranked on a composite of normalised citations at 40 percent, influential citations at 30, recency at 20 and relevance at 10. Nothing about how the shortlist was chosen is hidden.
Overrides that survive re-analysis
You know things the formula does not: that a registry entry was amended, that two cohorts double-count the same participants. Override a tier with a reason and the computed value stays on the record beside yours. Edited clusters and reports are pinned, so re-running the question refines your work instead of overwriting it.
Exports that cannot drift
Reports export to PDF, Markdown, JSON and HTML. The PDF is the print variant of the document on screen rather than a separate rendering path, so the copy you circulate and the copy you audit stay identical.
See Nodus on your own data.
Tell us what's slowing your business down.
A 30-minute conversation with an engineer, not a salesperson.