Paper-Agent
Paper-Agent 是一个面向科研人员和学生的智能论文检索与调研工具。项目基于多智能体协作架构(LangGraph),通过自然语言处理(NLP)、自动化搜索,帮助用户高效查找学术论
What is Paper-Agent?
Paper-Agent 是一个面向科研人员和学生的智能论文检索与调研工具。项目基于多智能体协作架构(LangGraph),通过自然语言处理(NLP)、自动化搜索,帮助用户高效查找学术论文、分析文献内容,并进行论文调研。Paper-Agent 支持多平台集成、关键词搜索、自动分析、论文调研,提升了学术研究的效率。适用于论文写作、学术调研、科研项目管理等多种场景,是学术调研的理想助手。
Overview of Paper-Agent
What it is
Paper-Agent is an intelligent paper retrieval and survey writing workbench designed for researchers and students. It is built on a multi-agent collaboration architecture (LangGraph) and uses natural language processing (NLP) and automated search to help users efficiently find academic papers, analyze literature, and write surveys. Version 2.0 is a complete rewrite of the earlier 1.x series, with a modern Vue 3 + TypeScript frontend, unified uv package management, expanded paper sources (arXiv, OpenAlex, Semantic Scholar), SQLite + filesystem session persistence, SSE-based real-time progress updates, and crash recovery.
Key features
- Multi-source paper retrieval: built-in connectors for arXiv, OpenAlex, and Semantic Scholar, unified as
PaperDocument, with filtering by year, source, count, and exclusion keywords, plus deduplication and scoring. - Progressive reading from abstract to full text: read abstracts first to judge relevance, then download, convert PDF to Markdown, chunk, extract, and write to a local vector store for qualifying papers; saves recovery state when dependencies are unavailable.
- Hierarchical research analysis:
AnalyseAgentfirst analyzes by subtopic per paper, then performs global synthesis to produce structured content on current status, consensus, controversies, gaps, temporal evolution, and outlook. - Evidence-constrained survey writing:
WritingOutlineAgentgenerates an outline with evidence mapping;WritingAgentwrites section by section, supplements evidence when insufficient, reviews, and revises up to a limited number of times. - Real-time session workbench: SSE pushes progress for retrieval, reading, analysis, outline, and section-by-section writing; SQLite + filesystem persistence preserves history across browser refreshes.
- Visual model configuration: manage provider protocols, API addresses, keys, three agent tiers, and embedding parameters in the browser; test connectivity with one click; changes take effect immediately.
Use cases
- Getting started with an unfamiliar research direction: automatically retrieve from multiple sources, deduplicate, filter, and organize papers by topic and constraints.
- Paper reading and note accumulation: read abstracts to judge relevance, then download full text, parse, chunk, and build a local searchable corpus.
- Writing a domain survey: generate a structured outline from subtopic and global analysis, then write section by section with evidence.
- Managing long-running tasks: session, run status, intermediate artifacts, and real-time progress are all visible in the workbench; can be saved and resumed at any time.
- Controlling model costs: use different model tiers for different stages and view actual token usage.
Getting started
- Install project dependencies in the project root:
powershell uv init uv venv --python 3.12 uv sync npm run front:install2. Create a local model configuration: copyconfig/model.example.jsontoconfig/model.jsonand ensure at least one provider withapi_base, a validapi_keyorapi_key_env, a configuredagents.default_agent, and a workingembedding_profiles.default_embedding. 3. Start the backend:powershell uv run python main.pyBackend listens on127.0.0.1:8000by default. 4. Start the frontend:powershell npm run front:devOpen the browser and go to the system settings page to test the model, then create a research task in the session workbench.
License and pricing
The license is not stated. The project is open source and free.
Based on the project's public README and repository metadata.