AI 产品案例 · 学习搜索 Harness

AI Compass

面向大学生学习场景的 AI 搜索 Harness：速学/深研双模式编排、多引擎聚合检索、生成式 UI，基于 Morphic 定制中文知识工作流。

已上线

01问题

大学生搜索要的不是事实，是"理解路径"：资料散落、引用不可信、中英文信息差——搜索结果和课堂知识之间缺一座桥。通用 AI 搜索给了答案，但给不了"按学习场景编排"的工作流。

02方案

构建学习场景的 Agent Harness：速学/深研双模式编排层（一次检索快速搞懂 vs 多轮交叉验证深挖），聚合 Tavily / SearXNG / Brave / Exa 多搜索引擎，Vercel AI SDK 5.0 流式渲染生成式 UI（卡片 / 表格 / 图片网格），内置课程检索、文档问答、AI 日报、GitHub 项目分析、毕设规划等中文知识工作流。

03我的职责

Harness 产品定位与双模式编排设计、中文知识工作流定制、多引擎检索层集成、Docker / Vercel 部署上线。

01 · 背景与痛点

普通搜索给"事实"，不给"怎么学"；学生需要的是带来源、可追溯、贴合中文学习习惯的答案。Harness 的价值是把"搜索-验证-理解"编排成可复用的流程。

02 · 产品设计与架构

方案：Morphic 基线 → 双模式编排（速学 / 深研）→ 多引擎聚合（Tavily / SearXNG / Brave / Exa）→ AI SDK 流式生成式 UI → 中文知识工作流。

我的职责：Harness 产品定位、双模式编排、中文工作流、多引擎集成、部署。

工作流设计

graph LR
A[学生提问] --> B[模式编排: 速学 / 深研]
B --> C[多引擎聚合检索]
C --> D[AI 整理 + 引用溯源]
D --> E[生成式 UI: 卡片 / 表格 / 图网格]

03 · 复盘

基于开源项目定制，最大的坑是上游升级会覆盖定制——把编排层、工作流、检索集成收敛为独立模块，并做升级演练；先跑通 Morphic 基线再逐层替换，改到一半再回头换基线几乎等于返工。

04 · 成果与产品思考

Vercel 上线，Docker 一键部署（PostgreSQL + Redis + SearXNG）；技术栈 Next.js 16 / React 19 / Vercel AI SDK 5.0 / Drizzle ORM / Bun；双语 README，Apache 2.0 开源。

产品思考：学习工具的价值是给"路径"，不是给"答案"；Harness 的护城河是编排，不是模型。

AI product case · Learning search harness

AI Compass

A learning-focused AI search harness for college students: dual-mode orchestration (speed-learn / deep-dive), multi-engine retrieval, and generative UI, customized on Morphic with Chinese knowledge workflows.

Live

01Problem

College students aren't searching for facts — they're searching for a path to understanding: scattered sources, untrustworthy citations, and a Chinese-English gap. Generic AI search returns answers, but not workflows orchestrated for learning scenarios.

02Approach

Built a learning-scenario Agent harness: a dual-mode orchestration layer (speed-learn for quick understanding vs deep-dive for multi-round cross-validated research), aggregating multiple engines (Tavily / SearXNG / Brave / Exa), with Vercel AI SDK 5.0 streaming a generative UI (cards / tables / image grids), plus Chinese knowledge workflows: course search, document Q&A, AI daily briefing, GitHub project analysis, and thesis planning.

03My role

Harness product positioning and dual-mode orchestration design, Chinese knowledge workflow customization, multi-engine retrieval integration, Docker / Vercel deployment.

01 · Background & Pain

Generic search returns facts, not a way to learn. Students need answers that are sourced, traceable, and tuned to how they actually study in Chinese. A harness turns "search → verify → understand" into a reusable flow.

02 · Product Design & Architecture

Approach: Morphic baseline → dual-mode orchestration (speed-learn / deep-dive) → multi-engine aggregation → AI SDK streaming generative UI → Chinese knowledge workflows.

My role: harness positioning, orchestration, Chinese workflows, engine integration, deployment.

Workflow

graph LR
A[Student question] --> B[Orchestration: speed-learn / deep-dive]
B --> C[Multi-engine retrieval]
C --> D[AI synthesis + citations]
D --> E[Generative UI: cards / tables / grids]

03 · Retrospective

Customizing an open-source project, the biggest trap is upstream upgrades overwriting your customizations — keep orchestration, workflows, and retrieval integration in isolated modules and rehearse upgrades. Get the Morphic baseline running first, then swap layers one at a time; rewiring mid-change is almost a full rework.

04 · Outcome & Product Thinking

Live on Vercel; Docker one-click deploy (PostgreSQL + Redis + SearXNG); stack: Next.js 16 / React 19 / Vercel AI SDK 5.0 / Drizzle ORM / Bun; bilingual README, Apache 2.0.

Product thinking: a learning tool's value is the path, not the answer — a harness's moat is orchestration, not the model.