AI Knowledge Graph：把任意文本扔进去，LLM 自动生成可交互知识图谱

by Mycelium Protocol

---

把一段非结构化文本喂给 LLM，能直接得到一张有意义的知识图谱吗？

AI Knowledge Graph（robert-mcdermott）给出了一个干净的答案：可以，而且是四阶段、全自动、输出可以直接在浏览器里交互的 HTML。2.7K stars，Apache-2.0，Python 3.11+，Ollama 本地就能跑。

---

它做的事

输入一个文本文件，输出一个可交互的知识图谱 HTML。中间的 LLM 可以是你本机的 Ollama，也可以是 OpenAI、vLLM、LM Studio、LiteLLM（后者还能桥接 AWS Bedrock、Azure OpenAI、Anthropic）——只要是 OpenAI 协议兼容的端点就行。

效果参考：工业革命知识图谱 demo（161 个节点，564 条边，9 个社区）。

---

四阶段处理流程

Phase 1：三元组提取

文本先被切成带重叠的 chunk（默认 200 词，20 词重叠），每个 chunk 单独送给 LLM，提取 Subject-Predicate-Object（主谓宾）三元组：

"瓦特改进了蒸汽机" → (瓦特, 改进了, 蒸汽机)
"蒸汽机推动了工业化" → (蒸汽机, 推动了, 工业化)

13 个 chunk，提取 216 条三元组——这是原始的、碎片化的关系网络。

Phase 2：实体标准化

同一个概念在不同 chunk 里可能有不同写法：“AI”、“人工智能”、“AI system”——在图里会是三个节点，连接断裂。

LLM 审阅所有唯一实体，识别同义词组，统一命名：

201 个原始实体 → 标准化为 181 个标准形式 → 最终 160 个唯一实体

这一步让图从碎片拼接变成连贯网络。

Phase 3：关系推断

即使标准化之后，图里还会有互相孤立的子社区——文本没有直接提到它们的关联。这一阶段做三件事：

1. Louvain 社区检测：识别图内的独立子图
2. LLM 推断跨社区关系：分析社区代表节点，推断合理的新关系
3. 传递推理 + 词法相似度推理：规则层的自动补全

原始 216 条三元组 → 推断补充 → 最终 564 条（增加 370 条）
实线 = 原始关系；虚线 = 推断关系

Phase 4：可视化

PyVis 生成单文件 HTML，内嵌 vis.js 物理引擎：

• 社区着色：Louvain 检测结果，不同社区不同颜色
• 节点大小：度中心性 + 介数中心性 + 特征向量中心性加权
• 实线/虚线：区分原始关系和推断关系
• 交互控制：缩放、平移、悬停查详情、物理引擎参数调节
• Light / Dark 主题

---

快速上手

git clone https://github.com/robert-mcdermott/ai-knowledge-graph
cd ai-knowledge-graph
uv sync

编辑 config.toml，接入本地 Ollama：

[llm]
model = "gemma3"
api_key = "sk-1234"
base_url = "http://localhost:11434/v1/chat/completions"
max_tokens = 8192
temperature = 0.2

[chunking]
chunk_size = 200
overlap = 20

[standardization]
enabled = true
use_llm_for_entities = true

[inference]
enabled = true
use_llm_for_inference = true
apply_transitive = true

运行：

uv run generate-graph.py --input your_text.txt --output graph.html

浏览器打开 graph.html，完成。

也可以作为包安装：

pip install --upgrade -e .
generate-graph --input your_text.txt --output graph.html

---

参数速查

参数	说明
--input FILE	输入文本文件
--output FILE	输出 HTML（默认 knowledge_graph.html）
--config FILE	配置文件路径（默认 config.toml）
--no-standardize	跳过实体标准化（更快，质量略低）
--no-inference	跳过关系推断（更省 token）
--debug	打印 LLM 原始响应，排查用
--test	用内置样例数据测试可视化

---

为什么值得关注

知识图谱构建一直是 NLP 里门槛比较高的方向——以前需要训练专门的 NER + 关系抽取模型，还要处理实体链接。LLM 的出现让 SPO 抽取变成了”写一个好的 prompt”，而这个项目把 prompt 工程、实体消歧、关系推断和可视化全部打包成了一个 uv run 能跑的工具。

兼容任意 OpenAI 协议端点意味着你可以完全本地（Ollama + Gemma3），也可以接 Claude/GPT——隐私敏感的文档不出本机，公开文档选最强模型，配置一行切换。

三阶段的处理设计（提取 → 标准化 → 推断）每一步都可以独立关闭，方便在速度和质量之间做取舍。

2.7K stars，382 forks，仍在活跃更新（最近更新 2026-08-04）。Apache-2.0 开源，可以直接嵌商业产品。

仓库：github.com/robert-mcdermott/ai-knowledge-graph · Demo：robert-mcdermott.github.io/ai-knowledge-graph

---

Mycelium Protocol — 追踪 AI 系统的底层演化

---

关于 Mycelium

菌丝协议。持续追踪 AI 工具、系统和实验的内容节点。

---

AI Knowledge Graph: Feed Any Text to an LLM, Get an Interactive Knowledge Graph

by Mycelium Protocol

Can you feed unstructured text to an LLM and get a meaningful knowledge graph out? AI Knowledge Graph (robert-mcdermott) answers that cleanly: yes — four phases, fully automated, output as interactive HTML you can explore in a browser. 2.7K stars, Apache-2.0, Python 3.11+, runs locally with Ollama.

What It Does

Input: a text file. Output: an interactive knowledge graph HTML file. The LLM in the middle can be local Ollama, OpenAI, vLLM, LM Studio, or LiteLLM (which bridges AWS Bedrock, Azure OpenAI, Anthropic, and more) — any OpenAI-protocol-compatible endpoint works.

See it live: Industrial Revolution knowledge graph demo — 161 nodes, 564 edges, 9 communities.

Four-Phase Pipeline

Phase 1: Triple Extraction

Text is split into overlapping chunks (default: 200 words with 20-word overlap). Each chunk goes to the LLM to extract Subject-Predicate-Object (SPO) triplets:

"Watt improved the steam engine" → (Watt, improved, steam engine)
"Steam engines drove industrialization" → (steam engine, drove, industrialization)

13 chunks, 216 extracted triplets — raw and fragmented at this point.

Phase 2: Entity Standardization

The same concept can appear multiple ways across chunks: “AI”, “artificial intelligence”, “AI system” — creating three separate nodes with broken connections. The LLM reviews all unique entities, identifies synonym groups, and unifies naming:

201 raw entities → standardized into 181 forms → 160 unique entities

This turns a fragmented patchwork into a coherent network.

Phase 3: Relationship Inference

Even after standardization, isolated sub-communities remain — text didn’t explicitly connect them. This phase does three things:

1. Louvain community detection: identifies disconnected sub-graphs
2. LLM-assisted cross-community inference: analyzes representative nodes and infers plausible new relationships
3. Transitive inference + lexical similarity rules: rule-based automatic completion

216 original triplets → inference adds 370 more → 564 total
Solid lines = original; dashed lines = inferred

Phase 4: Visualization

PyVis generates a self-contained HTML file with an embedded vis.js physics engine:

• Community coloring: Louvain results, distinct color per community
• Node size: weighted by degree, betweenness, and eigenvector centrality
• Solid vs. dashed lines: original vs. inferred relationships
• Interactive controls: zoom, pan, hover for details, physics controls
• Light / dark themes

Quick Start

git clone https://github.com/robert-mcdermott/ai-knowledge-graph
cd ai-knowledge-graph
uv sync

Configure config.toml for local Ollama:

[llm]
model = "gemma3"
api_key = "sk-1234"
base_url = "http://localhost:11434/v1/chat/completions"
temperature = 0.2

[standardization]
enabled = true
use_llm_for_entities = true

[inference]
enabled = true
use_llm_for_inference = true

Run:

uv run generate-graph.py --input your_text.txt --output graph.html

Open graph.html in a browser. Done.

Key Flags

Flag	Effect
--no-standardize	Skip entity standardization (faster, slightly lower quality)
--no-inference	Skip relationship inference (saves tokens)
--debug	Print raw LLM responses for debugging
--test	Test visualization with built-in sample data

Why This Matters

Knowledge graph construction used to require training dedicated NER + relation extraction models plus entity linking pipelines. LLMs make SPO extraction a prompt-engineering problem, and this project packages the prompt engineering, entity disambiguation, relationship inference, and visualization into a single uv run command.

Compatibility with any OpenAI-protocol endpoint means you can go fully local (Ollama + Gemma3) for sensitive documents or use the strongest available model for public content — one config line to switch.

The three-phase design (extract → standardize → infer) lets each step be disabled independently, making it easy to trade quality for speed or token budget.

2.7K stars, 382 forks, actively maintained (last update 2026-08-04). Apache-2.0, embeddable in commercial products.

Repository: github.com/robert-mcdermott/ai-knowledge-graph · Demo: robert-mcdermott.github.io/ai-knowledge-graph

---

Mycelium Protocol — tracking the deep evolution of AI systems

© 2026 Mycelium Protocol. All rights reserved.