hjcheng0602/paperwise

Paperwise, also called research-helper in its own files, is a command line tool that helps researchers read academic papers more deeply using AI language models. Rather than just asking an AI to summarize a paper in one shot, it runs a longer pipeline: it builds a vector based knowledge base to reduce made up answers, calls the language model multiple separate times for more depth, and tracks how ideas connect across papers in a knowledge graph. The main command reads a single paper, either by its arXiv ID or a local PDF file, and produces a structured report in six sections: the research problem and motivation, the core method including any formulas or architecture, the experiment design and results, a comparison with related work pulled automatically from the knowledge base, the paper's limitations and future directions, and a personal evaluation with research ideas it might inspire. Each of these six sections is generated by its own separate call to the language model rather than one combined request. Beyond single papers, the tool can search arXiv and produce a broader summary of a research area, search the built up knowledge base by meaning rather than exact keywords, and build an interactive knowledge graph from all the papers read so far. That graph connects concept nodes, like specific methods or techniques, to the papers that use them, and shows relationships such as one paper building on, comparing to, or contradicting another. The graph can be opened directly in a browser or exported for use in a separate graph visualization tool called Gephi. Setup involves cloning the repository, installing it with pip, and adding an API key to a configuration file. It supports several language model providers including DeepSeek, Qwen, OpenAI, and Anthropic, and can generate its own text embeddings locally without any API key if none of the cloud embedding providers are configured. The README estimates that reading one paper costs roughly 0.002 to 0.005 dollars in API fees. The project is released under the MIT license.