DataArc
联创 / 产品经理Co-founder & Product Manager
深圳 IDEA 研究院孵化Incubated at IDEA, Shenzhen
一家从合成数据起步的 AI 创业公司。我在团队形成早期加入，抽象共性需求，提出并推动 LivingKB 标品落地。An AI startup that began with synthetic data. I joined at an early stage, distilled recurring needs, and proposed and drove LivingKB toward a standard product.
01 团队与我的位置
Team & role
DataArc 是 IDEA 研究院孵化的合成数据团队。DataArc is a synthetic-data team incubated by IDEA.
我是初创期 PM，定义了早期产品 Roadmap。I was the founding-stage PM who defined the early product roadmap.
IDEA 研究院由沈向洋院士推动成立。DataArc 从 IDEA 的科研序列中孵化，沈向洋与郭健担任项目顾问，核心技术是语境图谱与合成数据：为稀缺、敏感和长尾场景生成可训练、可评测的数据。Academician Harry Shum helped establish IDEA. DataArc was incubated from IDEA's research organization, with Harry Shum and Jian Guo as project advisors. Its core technologies—Context Graphs and synthetic data—generate trainable and evaluable data for scarce, sensitive, and long-tail scenarios.
团队现在的产品与技术体系 Current product systemDataArc 官方网站DataArc official website新标签页打开 Open in a new tab
我的位置 My position团队刚形成时加入，承担早期共创角色，也是当时唯一的产品经理，后续负责组建产品与设计实习生团队。I joined as the team took shape, became its sole product manager in an early co-creation role, and later built the product and design intern team.
我从客户问题中提炼产品结构，组织方案版本与 POC，连接算法、工程、设计和商业材料。I distilled product structure from customer problems, organized proposals and POCs, and connected algorithms, engineering, design, and commercial material.
800 万种子轮计划8000 万估值 · 内部早期材料
RMB 8m seed plan · RMB 80m valuation
RMB 8m seed plan · RMB 80m valuation
三轮融资进程进入 2.5 亿投前阶段 · 内部汇报快照
Three-round trajectory · RMB 250m pre-money snapshot
Three-round trajectory · RMB 250m pre-money snapshot
数亿元投后估值连续完成种子、种子+轮 · 公开报道
Hundreds of millions post-money · Publicly reported
Hundreds of millions post-money · Publicly reported
IDEA · 官方 Official连续完成种子、种子+轮融资团队、技术、产品与融资Team, technology, products, and funding界面新闻 · 2025.11 Jiemian累计融资数千万元英诺天使、东方富海等参与Leading investors participated投中网 · 2025.11 ChinaVenture估值数亿的数据破壁者合成数据赛道与投资人判断Synthetic-data thesis
早期数字来自 2025 年内部融资材料；公开阶段采用 IDEA 与媒体发布口径。Early figures come from internal 2025 financing material; the public milestone follows IDEA and media releases.
02 产品判断
Product thesis
市场先听懂“知识库”The market understood “knowledge base” first
合成数据拥有明确的技术价值，但客户采购时更关心“你能为我做什么”。初创期，我们在短时间内接触几十家客户与合作方，从反复出现的问题里提炼出一个共性需求：企业知识库是 Agent 进入业务的前置环节，数据治理是业务 AI 化的第一步，知识治理在当时乃至今天都是重要刚需。Synthetic data had clear technical value, while buyers cared more about one question: “What can you do for my business?” After meeting dozens of clients and collaborators in a short period, we distilled a recurring need: enterprise knowledge bases precede agents, data governance is the first step toward operational AI, and knowledge governance remains a critical need today.
我们在短时间内接触了大量企业与合作方，从反复出现的业务问题中抽象出共性需求：企业知识库。In a short period, we met many enterprises and collaborators, then abstracted a shared need from recurring business problems: the enterprise knowledge base.
产品认知 Product insight合成数据是“性感”的融资叙事，也是底层的重要技术能力。但甲方更能听懂、且愿意买单的，是更贴近业务的“知识库”需求。Synthetic data makes a compelling fundraising story and remains an important foundational capability. Enterprise buyers, however, understood—and paid for—knowledge-base needs that connected directly to their business.
团队原有的三项核心能力Three core capabilities
合成数据Synthetic Data语境图谱Context Graph图上推理Think-on-GraphLivingKB主要标品 Core product
实际产品 Working product
LivingKB 产品演示LivingKB product demo
当时采用的产品形态类似 Manus 的多工具交互形态，底层是我们的 Living Knowledge Base。The product used a multi-tool interaction model similar to Manus, powered underneath by our Living Knowledge Base.
产品演示录屏 · 2025 Product demo · 2025LivingKB 是什么 What it is
持续生长的企业知识底座A living knowledge foundation
接入企业资料，建立语境图谱，让答案贴近业务知识库；在运行中动态合成新数据。It ingests enterprise material, builds a Context Graph, grounds answers in the business knowledge base, and dynamically synthesizes new data while running.
- 01整合知识Organize
- 02检索与推理Retrieve & reason
- 03持续扩充Expand
1业务交付
Delivery
Delivery
5POC 验证完成
Validated
Validated
4POC 进行中
In progress
In progress
2025 年 8 月业务快照。每个 POC 都由产品跟进，再把共性需求沉淀回 LivingKB 与 RAGFactory。August 2025 snapshot. Product followed each POC, then fed shared needs back into LivingKB and RAGFactory.
03 某教育公司 POC
Education case
POC：把标品带进具体业务，理解产品边界POC: Bring the standard product into a real workflow and understand its boundaries
该教育公司希望把教材、题库和知识点整理成可靠的知识体系，并进一步形成可采购的应用。过程概要：先梳理完整业务范围，再选择哲学作为最小切口；复用 LivingKB 与 Context Graph 的底座，同时为教育场景重新定义知识结构、数据处理和做题闭环。The education company wanted to organize textbooks, question banks, and concepts into a reliable knowledge system, then develop it into a purchasable application. We first mapped the full business scope, chose philosophy as the smallest viable wedge, reused the LivingKB and Context Graph foundation, and redesigned the knowledge structure, data processing, and practice loop for education.
客户目标 Client goal知识库 + AI 搜索 + 题目生成与练习Knowledge base, AI search, generation, and practice
复用底座 Reused foundationLivingKB + Context Graph + RAGFactoryIngestion, retrieval, graph reasoning, and evaluation foundations
场景定制 Domain work学科结构、教材解析、知识点关联与做题闭环Subject structure, parsing, concept links, and practice loop
让底层能力直接进入项目Carry the foundation into delivery
- 企业资料接入与解析 Knowledge ingestion
- 语境图谱建图与检索 Context Graph
- 答案溯源与图上推理 Traceability & reasoning
- 基础评测与算法框架 Evaluation foundation
围绕教育业务重新设计Design around the education workflow
- 从全量题库选择哲学切口 Choose the philosophy wedge
- 结构层、基础层、策略层 Three-layer knowledge model
- 教材与题目的解析规则 Textbook & question parsing
- 知识点—例题—练习闭环 Concept-to-practice loop
- 01理解业务 Understand
先看完整题库，再选最小切口Map the whole question bank, then choose the wedge
调研覆盖上线科目、二三级知识点、题量与需求。哲学兼具知识关联、适中难度和较小范围，适合在一周内跑通。Research covered subjects, topic hierarchies, volume, and demand. Philosophy combined relational knowledge, moderate difficulty, and a contained scope suitable for a one-week build.
- 02设计定制 Tailor
把行业知识写成产品结构Turn domain knowledge into product structure
教材目录形成结构层，教材知识点形成基础层，解题策略形成策略层；每一层都定义数据来源、关系和进入应用的方式。Textbook chapters formed the structure layer, concepts the foundation layer, and solving strategies the strategy layer—each with defined sources, relationships, and application paths.
- 03快速验证 Validate
一周跑通建库、检索与做题Complete the knowledge, retrieval, and practice loop in one week
四本教材约 750K tokens，1000 道题拆分建库与测试。评测用于检查复用底座在新场景里的速度和有效性。Four textbooks totaled about 750K tokens, with 1,000 questions split between graph building and testing. Evaluation checked the reused foundation's speed and effectiveness in the new domain.
< 6 minLivingKB 建图
Graph build
Graph build
110 minLightRAG 建图
Baseline build
Baseline build
< 2 s平均响应
Average response
Average response
64–66 81%初期准确率快照
Initial snapshots
Initial snapshots
这组数字只说明首轮 POC 的交付尺度：早期 PPT 与后续记录中的 LightRAG 基线分别为 64% 和 66%，LivingKB 初期结果均为 81%。These figures describe the scale of the first POC. The early deck and later note record LightRAG baselines of 64% and 66%, while both record an initial LivingKB result of 81%.
交付之后 After delivery
共性回到标品，领域逻辑留在应用Shared capabilities return to the product; domain logic stays with the application
语境图谱、检索与评测能力继续沉淀进 LivingKB 和 RAGFactory；该项目的学科分层、教材解析和做题逻辑服务于教育场景。让下一次交付拥有更成熟的通用底座，也保留面向业务重新设计的空间。Context Graph, retrieval, and evaluation capabilities continued to compound into LivingKB and RAGFactory. The subject hierarchy, textbook parsing, and practice logic remained specific to the education application, giving future projects a stronger shared foundation and room for business-specific design.
回看 Reflection
企业标品化的重要性Why standardization matters in enterprise products
DataArc 这段经历沉淀的经验：先从市场里找到共性问题，再把判断写进产品，用 POC 检查它，最后把交付中的共性能力沉淀回标品。My key takeaway from DataArc: identify recurring problems in the market, encode the product thesis, test it through POCs, and compound reusable capabilities from delivery back into the standard product.