A multi-agent pipeline that reads the papers and drafts the survey for you

What it does

You give it a research topic; a LangGraph pipeline of five agents (Search, Read, Analyse, Writing Outline, Writing) takes over. It pulls papers from arXiv, OpenAlex, and Semantic Scholar, deduplicates and scores them, reads abstracts to judge relevance, then downloads and chunks full texts into a local ChromaDB vector store. From there it builds a structured outline with evidence mapping and writes the survey section by section, retrieving more sources when evidence runs thin. A Vue 3 workbench streams every step’s progress over SSE, and SQLite persistence means a browser refresh won’t kill your session.

The interesting bit

The engineering discipline is the real story: each agent can run on a different model tier (cheap model for reading, stronger one for analysis), with actual token usage displayed. If a download dependency fails mid-run, it saves the recovery point and resumes later instead of dying — rare patience in this genre.

Key highlights

• Three-source retrieval unified into one PaperDocument model, with year/source/count/exclusion filters and dedup scoring
• Progressive reading: abstracts first, full-text processing only for papers that pass relevance
• Layered analysis — per-subtopic first, then global synthesis of consensus, disputes, and gaps
• Evidence-constrained writing with bounded review-and-revise loops
• OpenAI-compatible and Anthropic provider backends, configurable per agent in the browser

Caveats

• Requires you to bring your own LLM providers and API keys; nothing works without a configured default_agent
• The README doesn’t quantify output quality or report any benchmarks on the generated surveys
• 2.0 is a full rewrite of 1.x, so older workflows and docs may not carry over

Verdict

Worth a look if you regularly write literature surveys and want an auditable, resumable pipeline rather than a one-shot chatbot summary. If you just need quick paper summaries, it’s more machinery than you need.