# 实体库 · Graph — xiaopingfeng.com/graph/

> AI 行业里 App、模型、机构，以及它们之间带证据的关系（谁开发了谁、谁用着谁的模型、谁持有 / 收购 / 投资谁、哪个模型基于哪个模型）。
> 快照 2026-11-01：110 个 App、197 个模型、574 个机构，913 条关系；每条关系至少一条出处。
> 另有 46 个文件 / 规则（法律、条约、标准、案件、报告…）、27 个事件、41 个 Landscape 的层（`placed_in` / `flows_to` 把玩家和层挂在 Landscape 的结构图上，由页面派生）；本体见 https://xiaopingfeng.com/graph/data/ontology.ttl。
> 站点：https://xiaopingfeng.com/graph/ · 浏览：https://xiaopingfeng.com/graph/explore/ · 查询工作台：https://xiaopingfeng.com/graph/query/

## 读（不需要 key，已开 CORS）

- `GET https://xiaopingfeng.com/api/graph/query?type=model&where=access=open,eps=109&sort=-date&limit=20&format=md`
  参数：`type`（app | model | org | lab | neolab | inst | event | layer，逗号并集）· `where`（`字段=值`、`!=`、`~` 包含、`>=` `<=` `>` `<`，逗号分隔，值可写 `a|b`，`字段=*` 表示有值）·
  `of`（从某个实体出发，id 或名字）+ `rel` + `dir`（in | out | both）+ `hops`（1–3）· `sort`（`-date`）· `limit`（默认 100，最多 1000）· `group` ·
  `format`（json 默认 | md | csv | jsonl）· `fields`（json 时只留这些字段，如 `fields=name,date,developer`）
  「阿里开发的所有模型」：`/api/graph/query?of=Alibaba&rel=developed_by&dir=in&type=model`（developed_by 从产品指向开发者，所以找开发了什么用 dir=in）
- `GET https://xiaopingfeng.com/api/graph/entity/org:openai[?format=md]`  一个实体和它全部关系，每条关系带出处
- `GET https://xiaopingfeng.com/api/graph/search?q=black%20forest[&type=org]`  按名字 / 别名模糊找 id（写之前先查，避免重复）
- `GET https://xiaopingfeng.com/api/graph/schema`  实体、关系、出处类型的定义
- 静态数据：[kb.json](https://xiaopingfeng.com/graph/data/kb.json)（全部节点、关系、版面坐标、schema） · [kb.ttl](https://xiaopingfeng.com/graph/data/kb.ttl) · [kb.jsonld](https://xiaopingfeng.com/graph/data/kb.jsonld) · [kb-edges.csv](https://xiaopingfeng.com/graph/data/kb-edges.csv) · [schema.json](https://xiaopingfeng.com/graph/data/schema.json)

## 实体与 id

- `app:<slug>` App（来自 AppWatch 榜单，https://xiaopingfeng.com/live/appwatch/）
- `model:<id>` 模型（来自 AI 新模型台账，https://xiaopingfeng.com/landscape/ai-models/）
- `org:<id>` 机构；`profile` 字段指向本站的 NEO LAB / 实验室观察档案页
- 机构 = 主体（公司、实验室、非营利机构…），团队写在关系的 `unit` 上（如 Google DeepMind、通义 Qwen）

## 关系（每条都带 evidence：[{kind, label, text, url}]；每条都有方向：source → target）

下面每个关系类型后面的「读作」是把方向说成一句话的模板（`{s}` 是 source，`{t}` 是 target）；API 的 entity 结果里每条关系带 `sentence`，如「NVIDIA投资OpenAI」。

- `developed_by`  AIApp/Model → Organization（读作：{s}的开发方是{t}）：Points from the product to its developer. `unit` names the team inside the developer (e.g. Google DeepMind); `role` is primary or co.
- `uses_model_from`  AIApp → Organization（读作：{s}使用{t}的模型）：A documented dependency of an App on a lab's models. `weight` (optional) is how many of that lab's models.
- `uses_model`  AIApp → Model（读作：{s}使用模型{t}）：A documented dependency of an App on one specific model (an official doc or announcement names the model).
- `based_on`  Model → Model（读作：{s}基于{t}）：Source model = fine-tune / post-train / distillation of the target model. Evidence is the model card or announcement.
- `owns`  Organization → Organization（读作：{s}持有{t}）：Control or majority ownership. `since` / `until` bound it.
- `acquired`  Organization → Organization（读作：{s}收购{t}）：Completed or announced acquisition; `since` is the closing (or announcement) date, say which in `note`.
- `invested_in`  Organization → Organization（读作：{s}投资{t}）：Equity investment (round lead or named investor). Put the round and the amount in `note`; `since` is the announcement date.
- `supplies_compute_to`  Organization → Organization（读作：{s}向{t}供应算力）：A chip, cloud or data-center capacity supply agreement: source = supplier, target = customer. Put the scale (GW, chips, dollars) and the term in `note`; `since` is the announcement or signing date.
- `acquihired`  Organization → Organization（读作：{s}授权{t}的技术并挖走其团队）：Non-acquisition structure: the source licensed the target's technology and hired most of its team ('reverse acquihire'). Use `acquired` only for real acquisitions; if the parties reject the label, say so in `note`.
- `alumni_from`  Organization → Organization（读作：{s}的核心成员来自{t}）：The founders or core team of the source (typically a new lab) came from the target organization. Put the people (name, former role), and the unit when it is a team inside the target (e.g. Google DeepMind, Anthropic Alignment), in `note`/`unit`; `since` is when they left or the lab was announced. Evidence is the lab's own announcement or a primary profile — not a rumour about who might have left.
- `placed_in`  Organization/AIApp/Model/Instrument/Event → Layer（读作：{s}出现在 Landscape 的「{t}」）：Derived on every build, never written: an entity (or Instrument / Event) is the subject of a player card that stands on this layer of a Landscape map. Edge properties: cards (card ids), hi (the page highlights it), reported (every card for it is a `conflow` card: forecast, rumour or not yet settled). Evidence is the card on the site's own page.
- `flows_to`  Layer → Layer（读作：总览图里的箭头：{s} → {t}）：Derived on every build from the arrows of the overview diagram at the top of a Landscape map page (what the arrow means — upstream to downstream, cause to effect — is the map's own; see its page).
- `involves`  Instrument/Event → Organization/AIApp/Model（读作：{s}涉及{t}）：A document or an event (a lawsuit, a ruling, a statement, an incident, a regulatory action) and an organization, App or model it is about. `role` (required; several roles of one pair are joined with +, e.g. discloser+participant) says which part the target plays: plaintiff, defendant, regulator, issuer, speaker, subject, victim, attacker, discloser, vendor, customer, steward, participant, adjudicator (the court). Evidence is the Landscape card (or another source) that states it; `quote` in the evidence text is the sentence of the card.

出处类型：official, doc, filing, repo, modelcard, paper, own, media, wiki, social, other（另有建库时自动登记的 appstore、site、brand、a16z、play、ledger）。
机构和模型的 `landscape` 列出它出现在 Landscape 哪几张卡片（`role`：primary 是卡片的主角，mention 是被提到），Landscape 页面上每张卡也反过来链到这里。
`status`：缺省是基础表里的关系；`agent` = 智能体新增、还没人复核；`reviewed` = 已复核。机构的 `entry`：`base` | `agent` | `provisional`（由模型台账的开发者栏自动建立、待复核）。

## 每个实体一页

- `https://xiaopingfeng.com/graph/e/<org|app|model>/<slug>/`，例如 https://xiaopingfeng.com/graph/e/org/nvidia/ ；索引 https://xiaopingfeng.com/graph/e/ ；站点地图 https://xiaopingfeng.com/graph/sitemap.xml
- 页面里：带证据的关系（写成句子）、它在 Landscape 卡片 / NEO LAB 档案 / AppWatch / 新模型台账里的位置、哪几期 AI Buzzwords 提到过它（按名称匹配；数据在 [episode-mentions.json](https://xiaopingfeng.com/graph/data/episode-mentions.json)）。

- NEO LAB 实验室的关系图：https://xiaopingfeng.com/graph/neolab/（按研究方向着色，投资方与团队来源统计）；档案页仍在 https://xiaopingfeng.com/live/neolab/

## 嵌进网页

```html
<script src="https://xiaopingfeng.com/graph/embed.js" defer></script>
<xpf-query type="model" where="access=open,eps=109" view="table"></xpf-query>   <!-- list | table | cards | graph | count -->
<xpf-graph graph="kb" focus="Cursor" hops="2" height="420"></xpf-graph>
```

## 写

写入不开放给外部：条目由作者的智能体通过命令行工具追加（带校验、查重和出处要求），建库后随站点发布。发现错误或缺失，请在 https://xiaopingfeng.com/ 上联系作者。

## 可信度

关系的口径见各条出处；「待复核」的条目不要当事实引用。数据是 2026-11-01 的快照，关系会变。