设计了五层平台架构——Web App、项目工作区、Admin Dashboard、RAG Microservice 和 Local FAISS Service。
构建了统一的多提供商聊天层,通过单个会话线程,跨 OpenAI、Claude、Gemini、Grok 和 Perplexity 提供流式 SSE 响应。
提供了能力系统——由 Perplexity 驱动的带引用的网络搜索、深度研究会话、RAG 上下文注入的图像生成、文件上传 + RAG 索引,以及 Nexum Five 解决方案代理。
实现了 Split Mode 和 GrouPT——并排双窗格比较,以及带有辩论、头脑风暴和分析模式的结构化多 AI 讨论。
发布了 Artifacts 系统,支持 HTML/JSX/TSX 实时预览、版本历史、会话链接、社交功能和社区市场。
设计了工作流引擎(解决方案),支持多步自动执行、进度条、可共享链接和无需身份验证的访客执行。
构建了 Projects 协作工作区,结合了 TipTap 文档编辑器、Fortune Sheet 电子表格编辑器和工件代码编辑器——所有这些都基于项目文件进行 RAG 检索增强。
设计了双存储 RAG 架构——本地 FAISS 用于每个会话和每个 GPT 的知识,带有磁盘支持的索引;以及远程 Milvus 支持的 microservice,用于项目规模的工作负载。
实现了混合检索,结合了向量相似度、BM25 关键词匹配和新近度评分,并带有 IndexTag 范围搜索以实现个性化。
发布了 Admin Dashboard,具备用户管理、市场审核、按提供商/模型/用户成本分析、手动积分调整、定价配置和完整审计日志。集成了带有 5 层 RBAC 的 AWS Cognito 身份验证和基于 LemonSqueezy 的积分计费系统。
将五个领先的 AI 提供商统一到一个基于积分的界面下,增加了带有 RAG 基础的协作项目工作区,并引入了 Split Mode 比较和 GrouPT 小组讨论等多 AI 功能,这些功能是任何单个提供商原生不提供的。
MicrocosmWorks implemented a normalized credit system in NexumFive where each AI provider's token costs are mapped to a universal credit value maintained in the admin pricing configuration. When a user switches between OpenAI, Claude, Gemini, Grok, or Perplexity mid-conversation, the backend calculates credit consumption based on the per-model cost multiplier. The admin dashboard provides per-provider and per-model cost analytics for accurate margin tracking.
MicrocosmWorks designed the dual RAG system so local FAISS handles per-conversation and per-GPT knowledge with disk-backed indices for fast single-user retrieval, while the remote Milvus-backed microservice handles project-scale workloads requiring cross-document search across large file collections. Hybrid retrieval combines vector similarity, BM25 keyword matching, and recency scoring, with IndexTag-scoped search ensuring personalization within each user's data.
GrouPT is MicrocosmWorks' implementation of structured multi-AI discussions where multiple AI models participate in a conversation with defined roles. Users configure debate, brainstorm, or analysis modes, and the system orchestrates sequential responses from selected models (OpenAI, Claude, Gemini, Grok, Perplexity) via streaming SSE. Each model receives the full conversation context plus its role instructions, enabling comparative analysis of model strengths on the same prompt.
MicrocosmWorks built the Projects workspace combining TipTap for document editing, Fortune Sheet for spreadsheets, and an artifact code editor. All project files are automatically indexed into the Milvus RAG layer via a BullMQ embedding pipeline that processes documents on save. When users query the AI within a project context, retrieval is scoped to that project's indexed files, grounding responses in the team's actual documents and data.
MicrocosmWorks shipped a full marketplace system in NexumFive with community-published GPTs (custom AI personas with specialized instructions and knowledge), Workflows (multi-step automated processes with autopilot execution), and Artifacts (reusable HTML/JSX components with live preview). Users build custom GPTs through a no-code builder that configures system prompts, attached knowledge files indexed via FAISS, available tools, and model preferences.