Sedima
Getting started with Sedima.
Sedima is a research assistant built on top of a graph of claims, evidence, methods, and papers. This page is the short version — enough to try it in five minutes and understand what you're looking at.
Getting started
Open the workspace. Head to /workspace. Sign in with email or Google.
Ask your first question. Type a research question in the composer — plain English, no syntax. Example: "What are the main attention variants used in efficient transformers?"
Watch the trace. The panel on the right shows which papers the assistant retrieved, which claims it grounded on, and how it synthesized the answer. Every citation is clickable.
What is a claim graph
A claim graph is a representation of research literature where the primary objects are not papers but the individual claims those papers make, together with the evidence atoms that support each claim.
A typical paper contributes tens of claims — some empirical ("our method achieves X on Y"), some methodological ("we use technique Z"), some interpretive ("we hypothesize because A"). Each of those becomes a node. Each is anchored to the specific piece of the paper that supports it: a table row, a figure caption, a paragraph. The graph then connects claims across papers: agrees-with, refutes, extends, uses-same-benchmark, etc.
Why does this matter? Because the questions you want to ask at a literature-review scale — "who has shown X on Y?", "is there evidence against Z?", "what methods share this family?" — are graph queries, not document searches. Sedima does the extraction offline and lets you query the graph online.
How Novus works
Novus is the assistant that answers from the graph. Here is what happens when you hit Enter:
- Router decides whether your query needs the graph or is small talk. Trivial queries get an immediate reply. Substantive ones continue.
- Retrieval fans out across four retrievers in parallel: claim-level, paper-level, method-family, and citation-neighborhood. The winners are merged and packed by role.
- Synthesis streams an answer grounded in the packed sources. Every citation in the answer maps back to a specific evidence atom in the graph.
- Trace shows the full retrieval + synthesis timeline so you can audit the answer.
The full trace is always available under the workspace's Trace panel. Nothing is hidden.
Import your own paper
Sedima's public corpus is arXiv-first. To bring in a paper that is not already indexed, paste an arXiv ID or arXiv URL into the workspace composer. The pipeline runs:
arxiv-id → fetch → parse → extract atoms → build claims → link families → ready
Depending on load, the pipeline takes anywhere from thirty seconds (cached adapters) to a couple of minutes (fresh PDF that needs full parsing). Once done, the paper enters your library and Novus can answer from it.
What you can ask for
There is one composer and nothing to switch between. What used to be four buttons above it is four kinds of request, and Novus reads which one you meant:
- A question — the default. Answered from the graph, with the sources it used listed beside the answer.
- Verification — "where does this number actually come from?" follows one claim end-to-end down to the evidence atom. Ask it when an answer surprised you.
- Neighbours — "what else looks like this?" finds the method family a technique belongs to, and what sits next to it.
- Transfer — "this works in A, where else might it?" proposes hypotheses worth an experiment, each with the papers behind it.
Novus also works on your own drafts, not just the corpus: open one, read its LaTeX, propose an edit you accept or reject against a diff, or draw a figure into it from a template.