ResearchFlow is a local-first AI research automation prototype for LLM and agent research. It retrieves recent literature, records traceable evidence, proposes low-coverage research candidates, finds licensed datasets, exports isolated experiments, and creates citation-bound LaTeX manuscripts.
- Local registration and encrypted BYOK model profiles.
- OpenAlex, arXiv, Crossref, and Semantic Scholar aggregation with DOI/arXiv/title deduplication.
- Evidence extraction, coverage matrices, and ranked research-gap candidates with executed reverse searches, dated counterevidence, and recalibrated confidence.
- Hugging Face and OpenML dataset discovery, license gating, sample preparation, schema profiling, and byte-exact SHA-256 fingerprints.
- Model-generated or auditable fallback experiment code, AST safety checks,
uv.lock, and non-root Docker execution with no network, zero Linux capabilities, and strict resource limits. - Result-grounded arXiv projects plus official 2026 ICLR/ICML/NeurIPS template retrieval, LaTeX, BibTeX, PDF, template hashes, LLM disclosure, and claim provenance.
- Scientific-validity gates for dataset-topic fit, registered experiment plans, synthetic-label detection, post-run consistency, submission readiness, and sentence-level provenance.
- Evidence levels from concept draft through synthetic demonstration, initial experiment, reproducible research, and submission candidate.
- SSE progress updates, persistent workflow checkpoints, pause/resume, and model-call cost logs.
- SQLite/in-process mode plus PostgreSQL, Redis, Celery, pgvector semantic paper search, and the web application through Docker Compose.
Research-gap candidates mean low coverage within the retrieval snapshot. They never prove that no related publication exists.
中文新手教程：运行与使用指南 · Git 与 GitHub 版本管理 · 科学有效性等级
Requirements: Node.js 20+ and uv.
Copy-Item .env.example .env
uv sync --directory apps/api
npm install --prefix apps/web
.\scripts\start-local.ps1Open http://localhost:3000. The API documentation port follows API_PORT in .env.
Two visible process windows are opened. Close both windows to stop ResearchFlow.
The default storage root is D:\ResearchFlow\data. Before entering real API keys, replace
SECRET_KEY and ENCRYPTION_KEY in .env with long random values.
Run the elevated prerequisite installer once:
.\scripts\install-local-prerequisites.ps1It enables WSL2 and installs Docker Desktop. This workstation stores Docker's WSL virtual disks
under D:\ResearchFlow\docker\wsl; project data remains under D:\ResearchFlow\data.
Start PostgreSQL, Redis, the API, Celery worker, and web application:
docker compose up --buildWithout Docker, ResearchFlow will not execute generated code. It safely exports the experiment package instead.
In networks where Docker Hub is unavailable, set SANDBOX_BASE_IMAGE, API_BASE_IMAGE, and
WEB_BASE_IMAGE to trusted reachable mirrors. The current local .env uses the DaoCloud mirror
for the experiment sandbox; API and web mirrors can use the same registry prefix.
.\scripts\test-local.ps1