接入你的智能体
從發布者安裝。不是 PlugMCP 端點。
遠端 URL 還是本機指令？ 兩種安裝方式的差異
來自發布者
A semantic knowledge graph that extracts concepts from documents, tracks how well-supported they are, and remembers where sources disagree.
κ(G) — vertex connectivity of a graph. The minimum number of connections you'd need to cut before the graph falls apart. A measure of how robust the structure is.
Kappa Graph — κ(G). A semantic knowledge graph where knowledge has weight. Extracts concepts, measures grounding strength, preserves disagreement, traces everything to source.
Also kg — the unit of mass. Because knowledge here has weight. Grounding scores measure how heavy an idea is: well-evidenced claims carry more than thin ones. Contested concepts weigh differently than unchallenged ones.
The kg CLI, MCP server (for AI assistants), and optional FUSE filesystem:
curl -fsSL https://raw.githubusercontent.com/aaronsb/knowledge-graph-system/main/client-manager.sh | bash
Or just the CLI: npm install -g @aaronsb/kg-cli
Run your own knowledge graph backend:
curl -fsSL https://raw.githubusercontent.com/aaronsb/knowledge-graph-system/main/install.sh | bash
Or from source:
git clone https://github.com/aaronsb/knowledge-graph-system.git
cd knowledge-graph-system
./operator.sh init # Interactive setup
./operator.sh start # Start containers
See Quick Start Guide for details.
Interactive graph exploration with smart search, concept clustering, and relationship visualization
Command-line search returns concepts with source images rendered inline via chafa
t-SNE embedding landscape with auto-detected clusters, named by topic via TF-IDF
Ingest documents — PDFs, markdown, images, text. The system extracts concepts, relationships, and evidence automatically.
Search by meaning — "economic downturn" finds content about recessions, crashes, and crises even if those exact words aren't used.
Explore connections — Find paths between concepts. See how ideas relate across documents.
Check confidence — Every result includes grounding scores. Know what's well-supported vs. contested.
Trace sources — Every concept links back to the original text or image that generated it.
Query via AI — MCP server integration lets Claude and other assistants use the graph as persistent memory.
Navigate via filesystem — Mount the graph as a FUSE filesystem. Use ls, grep, find on semantic space.
Obsidian's graph view rendering knowledge graph relationships via the FUSE filesystem — no plugin needed
Research synthesis — Ingest papers, find connections across them, see where authors disagree. Grounding scores tell you which claims have broad support.
Technical documentation — Extract architecture concepts from diagrams, meeting notes, design docs. Query how components relate.
Agent memory — Give AI assistants persistent, grounded memory. They can check confidence before making claims.
Claude Desktop using MCP to search, explore relationships, and validate claims against the knowledge graph
Spatial understanding — Ingest place photos. The graph learns physical relationships without coordinates.
Compliance/audit — Full provenance chain. Every concept traces to source evidence.
Documents ──→ [FastAPI] ──→ LLM Extraction ──→ [PostgreSQL + AGE]
│ │
│ [graph_accel]
│ in-memory traversal
│ │
[Garage S3] [AGE graph store]
doc storage source of truth (ACID)
│ │
[React + D3] ←──── REST API ────→ [FastAPI]
web visualization query + ingest
│
[CLI / MCP / FUSE]
client interfaces
96 Architecture Decision Records document the design evolution.
Most systems that store knowledge for retrieval — vector databases, RAG pipelines, knowledge graphs — optimize for finding relevant content. They can tell you what matches your query. They can't tell you how well-supported it is, whether sources disagree about it, or where the evidence actually came from.
kg adds an epistemic layer on top of the graph. Every concept carries a grounding score computed from supporting vs. contradicting evidence — not a label, but a measurement. A concept backed by 47 sources with 12 contradictions scores differently than one with a single unchallenged mention. When sources disagree, the system preserves both sides rather than picking a winner.
Semantic diversity provides a second signal. Well-established knowledge tends to connect across independent domains. Narrow claims that only reference each other score lower. In testing, Apollo 11 mission data showed 37.7% diversity across 33 concepts; moon landing conspiracy content showed 23.2% across 3.
The system also handles images. Feed it street view photos and the extracted relationships ("next to", "across from", "visible from") naturally encode spatial topology — no coordinates needed.
Each neighbor leads on a different axis: Zep/Graphiti on bi-temporal fact tracking, GraphRAG on adoption (its Leiden communities are the closest analog to annealing, but computed once at index time), and Cognee — the nearest architectural neighbor — on self-improving memory (event-driven, no epistemic layer). kg's temporal axis is deliberately narrower: it records when evidence arrived (ingestion epoch), not when a fact was true — because it holds no truth values, only computed evidence, and preserves contradictions rather than expiring them. That's a stance, not a gap — see Computed Evidence over Asserted Truth. What we haven't found a shipped peer for is the combination: continuous ontology annealing, a semantic FUSE surface, and an epistemic layer that measures confidence rather than just retrieving content.
If you need:
kg was built for those requirements. Most alternatives optimize for retrieval accuracy or comprehensiveness. kg optimizes for knowing what you know and how well you know it.
Apache License 2.0 — Use, modify, distribute freely. Patent grant included.
Built with Apache AGE, Model Context Protocol, FastAPI, and local inference via Ollama.
Kappa Graph — κ(G). A semantic knowledge graph where knowledge has weight. Extracts concepts, measures grounding strength, preserves disagreement, traces everything to source.
工具
尚無工具清單。遠端伺服器會在首次健康檢查後取得工具清單。
常見問題
在本頁複製適用於 Cursor、Claude Desktop、Claude Code 或 VS Code 的 stdio 設定。它使用發布者的安裝方式，而不是 PlugMCP 代理伺服器。
這筆記錄未列出必要的環境變數。
Knowledge Graph System 透過 stdio 在本機執行。
上次檢查時間：2026-10-26。 Kappa Graph — κ(G). A semantic knowledge graph where knowledge has weight. Extracts concepts, measures grounding strength, preserves disagreement, traces everything to source.