tencent/wesmartflow

WeSmartFlow is an experimental learning platform from Tencent that uses AI agents to guide a student through a subject over time. The README, written mainly in Chinese with an English version linked, argues that most AI study tools are still chatbots that answer one question at a time, and that this project tries to model the wider learning process. The system tracks what the learner has covered, how well they grasp it, and what they should review next. The core of the product is a personal knowledge graph. As the learner talks with the AI tutor, new concepts are added as nodes, links between concepts are recorded, and a score for each node is updated on three axes: understanding, retention, and connection to other ideas. A spaced-repetition method called SM-2 decides when to bring a topic back for review. The same graph is shared across the interactive chat mode and a longer-form study mode, so progress is not lost between sessions. The tutor itself is built on a pattern called ReAct, where the model decides which tool to call next. The README lists the tools available to it: creating new graph nodes, updating mastery scores, generating PDF study cards through XeLaTeX and Beamer, producing quiz questions in four formats, searching the existing graph, fetching external information through Tavily, arXiv, or DuckDuckGo, and producing spoken explanations through the built-in macOS text-to-speech voice. There is also a multi-agent flow that turns a topic into a full course pack: a planning agent outlines chapters, a research agent gathers material, a writing agent produces slides, and other agents add images, audio, and exercises. The stack is FastAPI with SQLite on the backend and Vue 3 with Vite on the frontend. It needs Python 3.10 or newer, Node 18 or newer, a working LaTeX install for the PDF generation, and an OpenAI-compatible API key. The licence is MIT.