yongu.so is a desktop research assistant that turns a folder of PDFs into a thematically clustered, citation-backed LaTeX draft. It carries the mechanical load: extraction, deduplication, clustering, cite-keys, and cross-referencing. You keep every directional call, and every claim traces back to a specific sentence in a specific paper.
Runs on your machine. Your PDFs never leave it. Bring your own LLM key.
Eight capabilities already shipped in the first version. No freeform chat, just a deterministic state model you can actually trust.
Many independent projects, each its own local SQLite database. Nothing in the cloud.
Text extraction with best-effort metadata and a stable BibTeX cite-key generated at ingest.
Per-paper structured claims. On tool-capable models, a bounded loop cross-references your corpus.
Dedup and rewrite, then thematic clusters with a deterministic strength signal.
Support, contradiction, and methodology links, shown in an interactive claim graph.
Cluster synthesis into per-section LaTeX with inline \cite{}, exported ready to compile.
A provenance-backed Q&A surface over your claims. Answers cite the claims they came from.
Every artifact carries a passport: content hash, origin, lineage, freshness. An integrity check gates the draft step.
Data analysis, reference-code generation, cross-paper reasoning. The roadmap runs well past the literature review.
The system suggests the next step; you decide. It never auto-advances you or commits knowledge without an explicit action.
A single project state is the source of truth. The UI is a pure projection of it; events are the only thing that can change it.
Runs on your machine. SQLite and the filesystem hold everything. LLM calls go to a provider you configure.
The draft stage reproduces from the same clusters. Knowledge generation is auditable through recorded derivation traces.
AI tools hallucinate citations and you can't audit them. yongu.so doesn't. Every claim, cluster, and contradiction traces back to a specific sentence in a specific paper.
Native desktop binaries, built fresh on each release. Pick your platform.
Unsigned early build. On first launch, macOS may block it. Right-click the app → Open, or run xattr -d com.apple.quarantine yongu.
Unsigned early build. SmartScreen may warn. Click More info → Run anyway. No installer; it runs in place.
Unsigned early build. Extract and run: tar -xzf yongu-linux-x86_64.tar.gz && ./yongu. Needs GTK 3 and a GL-capable desktop (most distros have these).
~/yongu/llm_config.json or via an
OPENAI_API_KEY environment variable.
https://generativelanguage.googleapis.com/v1beta/openai, and use
gemini-3.1-flash-lite, the model this build is tuned for. It is fast and its free-tier rate limits are the most forgiving; a heavier Gemini model gives stronger synthesis but reaches the free quota sooner.
We're putting this in front of working researchers right now. What did you reach for that wasn't there? What would make the draft more useful? Would this fit your workflow? Every reply is read by the person building it.
Prefer email? me@yogi.sh