HackerNews中文版HackerNews CN
GitHubTwitter

我构建了一个151,000节点的GraphRAG群体，它能够自主发明可持续发展目标（SDG）解决方案。

2 分•作者: wisdomagi•2 个月前•原帖
嗨，HN， 我想分享一个我一直在构建的热情项目：PROMETHEUS AGI。 我对大多数大型语言模型（LLM）/检索增强生成（RAG）应用程序仅仅进行文本摘要感到沮丧。我想看看一个自主群体是否能够进行跨领域推理，从而发明新的物理解决方案（重点关注联合国可持续发展目标）。 技术栈： - Neo4j Aura（免费版已达到151,000个节点/400,000条边） - 数据获取：Google BigQuery（专利） + OpenAlex API - 大型语言模型：Ollama（Llama 3）用于零成本本地实体提取，Claude 3.5通过MCP进行深度推理。 - 用户界面：Streamlit（数字双胞胎仪表板） + React/Vite（着陆页）。 工作原理： 该群体将问题（例如，水过滤器中的生物污垢）映射到不同领域的孤立技术（例如，材料科学 + 纳米生物学），并寻找“缺失链接”——在专利数据库中尚不存在的组合。目前，该管道已自主草拟了261个以上的概念蓝图（如项目HYDRA，一个零功耗水净化器）。 我们目前正在寻找领域专家（工程师、材料科学家）来验证这些AI生成的蓝图并构建物理原型，同时也在寻求资金以将图谱扩展到超过100万个节点。 仪表板：https://project-prometheus-5mqgfvovduuufpp2hypxqo.streamlit.app/ 着陆页/演示文稿：https://prometheus-agi.tech 我非常希望听到你对架构、Neo4j模式设计或多代理方法的坦诚反馈！
查看原文
Hi HN, I wanted to share a passion project I've been building: PROMETHEUS AGI. I got frustrated that most LLM/RAG applications just summarize text. I wanted to see if an agentic swarm could actually perform cross-domain reasoning to invent new physical solutions (focusing on UN SDGs). The Stack: Neo4j Aura (Free tier maxed out at 151k nodes / 400k edges) Ingestion: Google BigQuery (Patents) + OpenAlex API LLMs: Ollama (Llama 3) for zero-cost local entity extraction, Claude 3.5 via MCP for deep reasoning. UI: Streamlit (Digital Twin dashboard) + React/Vite (Landing). How it works: The swarm maps problems (e.g., biofouling in water filters) to isolated technologies across different domains (e.g., materials science + nanobiology) and looks for "Missing Links"—combinations that don't exist in the patent database yet. So far, the pipeline has autonomously drafted 261+ concept blueprints (like Project HYDRA, a zero-power water purifier). We are currently looking for domain experts (engineers, materials scientists) to validate these AI-generated blueprints and build physical prototypes, as well as grants to scale the graph to 1M+ nodes. Dashboard: https://project-prometheus-5mqgfvovduuufpp2hypxqo.streamlit.app/ Landing/Deck: https://prometheus-agi.tech I would love to hear your brutally honest feedback on the architecture, the Neo4j schema design, or the multi-agent approach!