SciFork
SciFork — Git-native biomedical Research Graph. Literature-grounded, local-first, and auditable.
English | 简体中文
Fork hypotheses. Connect evidence. Advance research.
SciFork is a biomedical Research Graph plugin for DeepSeek Harness (DSH). DSH Chat stays your single conversation surface, while SciFork provides a visual Research Graph in the right sidebar to turn literature findings and discussions into structured, auditable research knowledge.
Current release: SciFork 0.0.8 targets DSH 0.2.1-rc.2 (desktop and same-version Web).
What you can do
- Deep research — ask for deep research in DSH Chat and SciFork runs a bounded multi-level exploration: each level retrieves literature, records machine-reviewed Evidence, keeps every qualifying branch, and automatically continues through the most promising new Hypothesis until your scope is covered.
- One-click expansion — select any node in the graph and click Research & Expand for a single grounded step outward (up to five direct candidate branches).
- Two-stage evidence gate — every Evidence assertion is bound to a PMID or DOI, a source locator, and a review rationale. Machine-reviewed Evidence can guide exploration, while only human-reviewed Evidence (or a validated team Result) can support an established Finding.
- Two views — see the full research structure in Main view or focus on a single node's direct assertions in Evidence view, with a Details panel for full Markdown notes and grouped citations.
- Plain files + local Git — your project lives in ordinary local Markdown and JSON files, and every change creates a local Git checkpoint on your current branch. You can connect the repository to GitHub or any remote of your choice.
Views and research model
- Main view (default) — displays the complete research structure: starting from open Research Questions, it connects candidate Hypotheses, testable Predictions, research-team Results, and established Findings, showing how claims support, contradict, cause, or predict one another.
- Evidence view — available whenever the selected node has direct literature support. It isolates that node and its linked Evidence assertions so you can inspect the underlying literature without visual clutter; switching back to Main restores your focus on that node.
- Details panel — shows the selected item's full Markdown body, publication and review counts, and citations grouped into human-reviewed, machine-reviewed, rejected, and retrieval references. Click the ID at the top of the panel to copy it when you want to reference a specific item in Chat.
Install and literature retrieval
From the Plugins panel (recommended): open Plugins in the DSH sidebar, select Add plugin, enter dsh-scifork, install, and click Enable now.
From a chat: ask the model — "install the DSH plugin dsh-scifork".
To uninstall, select Remove on the plugin's page; your local research project files remain intact and readable.
Pairing with literature retrieval
SciFork is not tied to a single database and works with real literature material in DSH Chat in this order:
- Retrieval plugin (recommended) — install
dsh-pubmed (tested with 0.4.4, optional) or another literature/PDF plugin so the model can search papers and read abstracts automatically.
- Web search — when no retrieval plugin is installed or results are insufficient, SciFork falls back to DSH's built-in web search.
- Pasted source text — you can also paste abstracts or bounded PDF passages directly into Chat along with a valid PMID or DOI.
Across all paths, only records with a valid PMID or DOI and real abstract or passage text are accepted; title-only records are skipped and citations are never fabricated.
Quick start
Choose an empty directory or the root of a Git repository (not a subdirectory inside another Git repo, and make sure Git user.name and user.email are configured). Initialize the directory once in DSH Chat:
/research init
Then open Research Graph from the right-sidebar guide card or the Research Graph button at the bottom of the left sidebar:
- Ask a research question — describe an open biomedical question in DSH Chat; SciFork records it as your starting Research Question and focuses it in the graph.
- Retrieve and expand — ask DSH to retrieve literature and import evidence; select a node and click Research & Expand for a single step, or ask in Chat for deep research to explore across multiple levels.
- Review and promote — switch between Main and Evidence views to inspect claims and sources, then ask in Chat to accept verified machine-reviewed assertions as human-reviewed Evidence (or critique, edit, or delete weak branches) to advance a Hypothesis into a Finding.
- Check project health — validate your project at any time in Chat:
/research validate
The Research Graph tab is a read-only surface for browsing, focusing nodes, and triggering one-step expansion. Day-to-day questions, deep research, edits, branch deletion, and Git history recovery all happen in DSH Chat.
Data and safety
SciFork runs locally and never uploads your research data or adds cloud sync. Project files store only PMIDs/DOIs, article titles, journals, years, and derived assertions with locators — full abstracts and PDFs are never written into your repository.
Before committing or sharing a project, check that it does not contain protected health information (PHI/PII) or controlled-access data. See SECURITY.md for details.
License
SciFork is available under the MIT License. Your research project data remains yours and may carry its own ownership and sharing terms.
