JFox

A local-first Zettelkasten knowledge management CLI. Bidirectional links, semantic search, knowledge graphs — all offline, all on CPU.

JFox (J + Fox / "box") is a command-line tool that helps you build a personal knowledge base using the Zettelkasten method. Notes live as plain Markdown files on your disk, connected by [[wiki links]] and indexed for instant semantic search.

Table of Contents

• Features
• Architecture
• Data Flows
• Quick Start
• Command Reference
• Note Format
• Contributing
• License

Features

• Three note types — Fleeting (quick capture), Literature (reading notes), Permanent (refined knowledge)
• Bidirectional links — Write [[Note Title]] to connect notes; backlinks are auto-generated
• Hybrid search — BM25 keyword search + semantic vector search, fused with Reciprocal Rank Fusion
• Knowledge graph — NetworkX-powered link analysis: clusters, orphans, hubs, shortest paths
• File watcher — Real-time index updates when you edit notes with any editor
• Multi knowledge bases — Manage separate KBs for work, personal, research, etc.

Architecture

Three-Layer Design

graph TB
    subgraph CLI ["CLI Layer"]
        cmd["jfox commands<br/>(Typer)"]
    end
    subgraph Storage ["Storage Layer"]
        note[note.py]
        models[models.py]
        md[("Markdown Files<br/>YAML Frontmatter")]
    end
    subgraph Index ["Index Layer"]
        se[search_engine.py<br/>HybridSearchEngine]
        vs[vector_store.py<br/>ChromaDB]
        bm[bm25_index.py<br/>BM25Okapi]
        emb[embedding_backend.py<br/>all-MiniLM-L6-v2]
        daemon["daemon/<br/>HTTP Server"]
    end
    subgraph Analysis ["Analysis Layer"]
        gph["graph.py<br/>NetworkX DiGraph"]
    end
    subgraph Watcher ["File Watcher"]
        idx[indexer.py<br/>watchdog]
    end

    cmd --> note & se & gph
    note --> models --> md
    se --> vs & bm
    vs --> emb
    emb -.->|"preferred"| daemon
    idx --> vs

Module Map

Data Flows

Note Creation

When you run jfox add, the system parses wiki links, creates the Markdown file, updates both indexes, and propagates backlinks:

sequenceDiagram
    participant U as User
    participant CLI as cli.py
    participant NM as note.py
    participant MD as Filesystem
    participant VS as VectorStore
    participant BM as BM25Index
    participant T as Target Note

    U->>CLI: jfox add "content with [[Link]]"
    CLI->>CLI: extract_wiki_links() → ["Link"]
    CLI->>CLI: find_note_id_by_title_or_id()
    Note over CLI: Match: exact ID → exact title → substring
    CLI->>NM: create_note(content, links=[id1])
    NM->>NM: generate_id() → timestamp + random
    NM->>MD: write Markdown + YAML frontmatter
    NM->>VS: add_note() → embed + store in ChromaDB
    NM->>BM: add_document() → tokenize + update index
    CLI->>T: load target → append backlink → save

Index Rebuild

jfox index rebuild reconstructs both the vector index and the keyword index from all Markdown files on disk:

sequenceDiagram
    participant U as User
    participant CLI as cli.py
    participant IDX as Indexer
    participant VS as VectorStore
    participant BM as BM25Index
    participant FS as Filesystem

    U->>CLI: jfox index rebuild
    CLI->>VS: clear() — wipe ChromaDB collection
    CLI->>FS: rglob("*.md") — scan all notes
    loop Each note file
        FS-->>IDX: parse Markdown + frontmatter
        IDX->>VS: add_or_update_note()
        Note over VS: embed → store in ChromaDB
    end
    CLI->>FS: list all notes
    CLI->>BM: rebuild_from_notes()
    Note over BM: tokenize all → rebuild BM25Okapi → persist

Hybrid Search (BM25 + Semantic → RRF)

jfox search runs two independent search paths in parallel and fuses results using Reciprocal Rank Fusion:

sequenceDiagram
    participant U as User
    participant SE as HybridSearchEngine
    participant VS as VectorStore
    participant BM as BM25Index
    participant EMB as EmbeddingBackend

    U->>SE: search("knowledge management", mode=hybrid)
    par Semantic Path
        SE->>EMB: encode(query) → 384-dim vector
        EMB-->>VS: cosine similarity search
        VS-->>SE: ranked results with scores
    and Keyword Path
        SE->>BM: tokenize(query) → BM25 scoring
        BM-->>SE: ranked results with scores
    end
    Note over SE: Graceful fallback if one path fails
    SE->>SE: RRF Fusion: score = Σ 1/(k + rank), k=60
    SE-->>U: merged, re-ranked results

Query with Graph Traversal

jfox query combines hybrid search with knowledge graph BFS to find semantically related notes and their neighbors:

sequenceDiagram
    participant U as User
    participant CLI as cli.py
    participant SE as SearchEngine
    participant KG as KnowledgeGraph
    participant NX as NetworkX

    U->>CLI: jfox query "Luhmann's methodology" --depth 2
    CLI->>SE: hybrid search → top results
    SE-->>CLI: ranked search results
    CLI->>KG: build() — 3-pass graph construction
    Note over KG: Pass 1: nodes from files<br/>Pass 2: edges from frontmatter links<br/>Pass 3: edges from [[wiki links]]
    loop For each search result
        CLI->>KG: get_related(note_id, depth=2)
        KG->>NX: BFS traversal (predecessors + successors)
        NX-->>KG: neighbors grouped by depth
    end
    CLI-->>U: results enriched with graph context

Quick Start

Install

# Recommended
uv tool install "git+https://github.com/zhuxixi/jfox.git"

# Or with pip
pip install -e ".[dev]"

See Installation Guide for details, Windows PATH setup, and uninstall instructions.

Create Your First Note

jfox init
jfox add "The Zettelkasten method uses atomic notes connected by links" \
    --title "Zettelkasten Introduction" --type permanent

Add Links

jfox add "[[Zettelkasten Introduction]] was invented by Niklas Luhmann" \
    --title "Luhmann and the Card Box" --type permanent

The [[Zettelkasten Introduction]] syntax automatically creates a bidirectional link. Backlinks are propagated to the target note.

Search

# Semantic + keyword hybrid search
jfox search "knowledge management method"

# Hybrid search + graph traversal
jfox query "Luhmann's methodology" --depth 2

Command Reference

Knowledge Base

Notes

Search & Analysis

Index Management

Templates

Performance & Debug

Daemon

Auto-Summary

Auto-summary runs inside the daemon to automatically archive Claude Code sessions into your knowledge base. It scans ~/.claude/projects/ for finished sessions, generates a structured summary via claude -p, and writes it as a session type note.

A session is considered "finished" when its file has not been modified for idle_threshold minutes (default: 30).

Key options:

• --interval — Scan interval in minutes (default: 30)
• --idle-threshold — Minutes of inactivity to consider a session finished (default: 30)
• --kb — Target knowledge base for saved notes

How it works:

• Uses claude -p (non-interactive mode) to generate summaries from stdin/stdout
• Runs with --permission-mode bypassPermissions so the daemon never blocks on permission prompts
• Tracks processed sessions in ~/.zk_auto_summary_state.json to avoid duplicates; transient failures are retried up to 3 times before giving up
• A session is eligible only after its file has been idle for idle_threshold minutes (no new content)

Privacy note: Auto-summary sends session text to Anthropic API via claude -p to generate summaries. Only session content is transmitted.

Self-Update

Global Options

Agent Plugins

JFox 提供主流 AI Agent 的插件/技能集成：

Claude Code

• 插件目录：packages/cc-plugin/
• 安装（marketplace 上架后）：/plugin marketplace add zhuxixi/jfox
• 包含 5 个 skill：manage、search、ingest、organize、session-summary

Kimi Code CLI

• 插件目录：packages/kimi-plugin/
• 安装：在 Kimi Code CLI TUI 中执行 /plugins install github:zhuxixi/jfox?path=packages/kimi-plugin，或使用本地 zip
• 包含 6 个 skill：using-jfox（会话启动自动加载）、jfox-manage、jfox-search、jfox-ingest、jfox-organize、jfox-session-summary
• 详见：packages/kimi-plugin/README.md

其他 Agent

• skills-recommend/pi/ 提供 pi coding agent 的技能包
• skills-recommend/kimi-cli/ 保留旧版 Kimi CLI 手动复制 skill（已弃用）

Note Format

Directory Structure

~/.zettelkasten/
├── default/                # Default knowledge base
│   ├── notes/
│   │   ├── fleeting/       # Quick captures
│   │   ├── literature/     # Reading notes
│   │   └── permanent/      # Refined knowledge
│   └── .zk/
│       ├── chroma_db/      # Vector index
│       ├── bm25_index.pkl  # Keyword index
│       ├── templates/      # Jinja2 templates
│       └── config.yaml     # KB config
├── work/                   # Named KB example
│   ├── notes/
│   └── .zk/
└── ~/.zk_config.json       # Global KB registry

File Format

Each note is a Markdown file with YAML frontmatter:

---
id: '20260321011528'
title: Machine Learning Overview
type: permanent
created: '2026-03-21T01:15:28'
updated: '2026-03-21T01:15:28'
tags:
- ml
- ai
links:
- 20260321011546
backlinks:
- 20260321011550
---

# Machine Learning Overview

[[Deep Learning]] is a subfield of machine learning...

Note Types

Link Resolution

[[Link Text]] matches notes by priority:

1. Exact ID — if text matches a note ID
2. Exact title — case-insensitive title match
3. Substring — title contains the link text

Contributing

git clone https://github.com/zhuxixi/jfox.git
cd jfox
uv sync --extra dev
uv run pytest tests/ -v

See Troubleshooting for common issues.

License

MIT

Acknowledgments

• sentence-transformers — text embeddings
• ChromaDB — vector database
• NetworkX — graph algorithms
• Typer — CLI framework
• Rich — terminal formatting