rbrain — Base de Conhecimento Pessoal com IA
hongyuatcufe/rbrain
Base de conhecimento pessoal + busca híbrida + grafo conceitual
Ver no GitHubServidor MCP em Rust que implementa uma base de conhecimento pessoal para pesquisa acadêmica, com busca híbrida (BM25 + vetorial), grafo de conhecimento, extração automática de conceitos e suporte nativo para CJK. Opera como binário único independente com SQLite, tantivy e usearch.
Instalação
npx convez-ai hongyuatcufe/rbrain1 comando no terminal: detecta o formato do repositório (marketplace ou plugin avulso) e instala via Claude Code. Requer Node 18+ e o Claude Code instalado.
Métricas
Casos de uso
- Organizar e conectar literatura académica com evidências específicas
- Busca semântica + full-text em corpus de pesquisa multilíngue
- Extração automática de conceitos e síntese de clusters conceituais
- Integração com Claude para raciocínio estruturado sobre contexto recuperado
Veredito

Ferramenta robusta e bem arquitetada para pesquisadores que precisam gerenciar corpora complexos multilingues. A integração MCP com Claude é valiosa, mas exige configuração de APIs externas (embeddings, LLM) e curva de aprendizado significativa em Rust. Implementação de qualidade produção para usuários avançados.
README
A tradução para o português deste README ainda está na fila.
rbrain
rbrain is a Rust-based personal AI knowledge base CLI for academic research. A self-contained binary backed by SQLite, tantivy (BM25 full-text search), and usearch (vector search) — no system dependencies required.
rbrain 是面向学术研究的个人 AI 知识库命令行工具。单一自包含二进制,后端使用 SQLite + tantivy(BM25 全文检索)+ usearch(向量检索),无任何系统依赖。
Features / 功能特性
- Hybrid search / 混合检索 — BM25 keyword search + semantic vector search, optional LLM query expansion BM25 关键词检索 + 语义向量检索,支持 LLM 查询扩展
- Knowledge graph / 知识图谱 — typed directed links anchored to specific passages (
evidence,related,supports,contrasts,develops) 带类型的有向链接,可锚定到具体段落 - Dream Cycle / 自动知识提取流水线 — lint → embed → extract concepts/figures → synthesize concept clusters 自动化多阶段流水线:检查 → 嵌入 → 提取概念/人物 → 概念聚类综合
- Think / 深度推理 — structured reasoning over retrieved context: tensions, judgments, open questions 基于检索上下文的结构化推理:张力、工作判断、开放问题
- Timeline / 时间线 — dated evidence log attached to any page 为任意页面附加带日期的证据条目
- Takes / 诠释片段 — interpretive fragments (
judgment,question,hypothesis,interpretation) without overwriting page content 为页面附加诠释性笔记,不覆盖原文 - MCP server / MCP 服务端 — expose all brain operations to Claude and other MCP-compatible AI assistants 将所有知识库操作暴露给 Claude 等 MCP 兼容客户端
- CJK support / 中日韩支持 — language detection (zh-hans, zh-hant, ja, ko, en), CJK-safe chunking, mixed-script search 语言检测、CJK 安全分块、多语言混合检索
Install / 安装
git clone https://github.com/hongyuatcufe/rbrain
cd rbrain
cargo build --release -p rbrain-cli
# add to PATH
ln -sf "$PWD/target/release/rbrain" ~/.local/bin/rbrain
Requires Rust 1.78+. / 需要 Rust 1.78+。
Quick Start / 快速上手
# Initialize a project brain / 初始化项目知识库
cd my-research-project
rbrain init
# Configure API keys (.rbrain/config.toml is gitignored)
# 配置 API 密钥(.rbrain/config.toml 默认已加入 .gitignore)
cat > .rbrain/config.toml <<EOF
[qwen]
api_key = "sk-..."
base_url = "https://dashscope.aliyuncs.com/compatible-mode/v1"
model = "text-embedding-v4"
[deepseek]
api_key = "sk-..."
base_url = "https://api.deepseek.com/v1"
model = "deepseek-chat"
embedding_dim = 1024
EOF
# Import source articles / 导入源文献
rbrain import ./papers/
# Run full dream cycle / 运行完整自动化流水线
# embed → extract concepts/figures → synthesize
rbrain dream
# Search and retrieve / 检索
rbrain search "自主知识体系"
rbrain query "中国教育学的本土化路径" --expand
rbrain get concepts/自主知识体系
CLI Reference / 命令参考
Brain Management / 知识库管理
rbrain init # create .rbrain/ in CWD / 在当前目录初始化
rbrain stats # page/chunk/link/embedding counts / 统计信息
rbrain doctor # health check / 健康检查
rbrain lint # issues: broken links, orphans, unembedded pages
Reading & Writing Pages / 读写页面
rbrain put <slug> --file <path> # write page from file / 从文件写入页面
rbrain get <slug> # read page + timeline / 读取页面及时间线
rbrain list [--type <type>] [--tag <t>] # list pages / 列出页面
rbrain import <dir> # import all .md files / 批量导入
rbrain export --dir <out> --format md # export as Markdown / 导出为 Markdown
rbrain export --dir <out> --format json # export as JSON / 导出为 JSON
Page types / 页面类型: note | concept | figure | synthesis | wiki | question | evidence | draft | memo | period | book
Search & Retrieval / 检索
rbrain search "query" # BM25 keyword search / BM25 关键词检索
rbrain search "query" --tag t --type concept # filtered / 过滤检索
rbrain query "question" --expand # hybrid + query expansion / 混合检索
Knowledge Graph / 知识图谱
rbrain link <from> <to> --type evidence --from-chunk <id> # 链接到具体段落
rbrain link <from> <to> --type related
rbrain unlink <from> <to> [--type <t>]
rbrain backlinks <slug> # incoming links / 反向链接
rbrain links <slug> # outgoing links / 出向链接
rbrain graph-query <slug> --depth 2 --direction both
rbrain orphans # pages with no incoming links / 孤立页面
rbrain extract --all # extract [[wikilinks]] from content / 提取双链
Link types / 链接类型: evidence | related | supports | contrasts | develops
Timeline / 时间线
rbrain timeline <slug> --text "..." --date "2026-05-15" --source "<slug> chunk:<id>"
# --date defaults to today / --date 默认为今天
Takes / 诠释片段
rbrain take <slug> "judgment text" --kind judgment
rbrain takes <slug> # list all takes / 列出所有诠释
--kind: judgment | question | hypothesis | interpretation
Tags / 标签
rbrain tag <slug> <tag>
rbrain untag <slug> <tag>
rbrain tags <slug>
Synthesis / 综合分析
rbrain think "topic" --limit 12 --expand # deep reasoning / 深度推理
rbrain think "topic" --save # saves as synthesis/<slug>
rbrain generate "topic" --limit 12 --expand # wiki-style summary / 文献综述草稿
rbrain generate "topic" --save # saves as wiki/<slug>
Embeddings / 向量嵌入
rbrain embed --stale # re-embed stale pages / 重新嵌入过期页面
rbrain embed --all # re-embed everything / 重新嵌入全部
rbrain sync --embed # sync filesystem → DB then re-embed / 同步后重新嵌入
Dream Cycle / 自动化流水线
rbrain dream # full pipeline / 完整流水线
rbrain dream --stage lint # Phase 1: lint
rbrain dream --stage embed # Phase 2: embed stale pages / 嵌入过期页面
rbrain dream --stage extract # Phase 3: extract concepts & figures / 提取概念和人物
rbrain dream --stage synthesize # Phase 4: synthesize concept clusters / 概念聚类综合
Phase 3 — Extract / 提取阶段:对每篇未处理的 note,调用 DeepSeek 提取概念、学者/人物和时间线事件。自动创建 concepts/<slug> 和 figures/<slug> 页面;关联人物的事件写入人物页,其他事件写入 research/evidence/events/<source-slug> 派生页。raw/ 来源文献不会被 dream 改写。处理记录写入 dream_metadata(幂等,不重复处理)。
For each unprocessed note, calls DeepSeek to extract concepts, figures, and timeline events. Events associated with people are written to figure pages; other events are written to derived research/evidence/events/<source-slug> pages. Dream never rewrites raw/ source documents. Idempotent via dream_metadata.
Phase 4 — Synthesize / 综合阶段:对有 3 篇以上源文章反向链接的概念,自动生成结构化文献综合页面,保存至 synthesis/<concept-slug>,并建立 develops(→ 概念)和 evidence(→ 源文章)链接。源文章更新后自动重新综合。
For each concept with 3+ source note backlinks, generates a structured literature synthesis at synthesis/<concept-slug>. Auto-links synthesis → concept (develops) and synthesis → source notes (evidence). Re-synthesizes when sources are updated.
MCP Server / MCP 服务端
rbrain 通过 Model Context Protocol 将所有操作暴露给 Claude Code 等 MCP 客户端。
rbrain serve mcp # stdio mode (for Claude Code) / stdio 模式
rbrain serve mcp --http 127.0.0.1:3456 # local HTTP mode / 本地 HTTP 模式
rbrain serve mcp --http 0.0.0.0:3456 --allow-remote # explicit remote bind / 显式远程绑定
HTTP mode exposes mutation tools without authentication and therefore only accepts loopback
addresses (127.0.0.1, [::1], or localhost) by default. Use --allow-remote only behind
a trusted boundary such as a local tunnel, firewall, or reverse proxy with authentication.
MCP Tools / MCP 工具列表
| Tool | Description |
|---|---|
brain_put | Write or update a page / 写入或更新页面 |
brain_get | Read a page / 读取页面 |
brain_delete | Delete a page / 删除页面 |
brain_list | List pages with optional type/tag filter / 列出页面 |
brain_query | Hybrid semantic search / 混合语义检索 |
brain_think | Deep reasoning synthesis on a topic / 深度推理 |
brain_generate | Search + LLM wiki synthesis / 文献综述生成 |
brain_link | Create a typed graph link / 建立知识图谱链接 |
brain_unlink | Remove a link / 删除链接 |
brain_backlinks | Get incoming links / 反向链接 |
brain_outlinks | Get outgoing links / 出向链接 |
brain_graph | Traverse graph neighborhood / 图谱邻域遍历 |
brain_orphans | List pages with no incoming links / 孤立页面 |
brain_add_timeline_entry | Add a dated entry to a page / 添加时间线条目 |
brain_add_tag | Add a tag / 添加标签 |
brain_remove_tag | Remove a tag / 删除标签 |
brain_stats | Brain statistics / 统计信息 |
Claude Code Setup / Claude Code 配置
Add to .claude/settings.json / 添加到 .claude/settings.json:
{
"mcpServers": {
"rbrain": {
"type": "stdio",
"command": "rbrain",
"args": ["serve", "mcp"]
}
}
}
Architecture / 架构
rbrain/
├── crates/
│ ├── rbrain-cli/ # CLI entry point (clap) / 命令行入口
│ ├── rbrain-engine/ # Core logic: dream, search, linking, synthesis / 核心逻辑
│ ├── rbrain-core/ # Page model, language detection, markdown parsing / 页面模型
│ ├── rbrain-db/ # SQLite schema and queries (sqlx) / 数据库层
│ ├── rbrain-search/ # Tantivy BM25 + usearch vector index / 检索层
│ ├── rbrain-llm/ # DeepSeek chat + Qwen embedding clients / LLM 客户端
│ ├── rbrain-mcp/ # MCP server (stdio + HTTP) / MCP 服务端
│ └── rbrain-worker/ # Background job queue / 后台任务队列
└── rbrain-research-cli/ # Claude Code skill for research workflows / 研究工作流技能
Embedding / 嵌入模型: Qwen text-embedding-v4 (1024-dim) via DashScope
LLM: DeepSeek deepseek-chat — extraction, synthesis, think, generate
Storage / 存储: SQLite (pages, chunks, links, dream_metadata) + usearch (vectors) + tantivy (BM25)
Brain Auto-Discovery / 知识库自动发现
rbrain 从当前目录向上遍历自动定位知识库:
- 找到
.rbrain/→ 使用该项目本地知识库 - 未找到 → 回退到
~/.rbrain/
rbrain walks up from CWD to discover the active brain:
- finds
.rbrain/→ uses project-local brain - falls back to
~/.rbrain/
这意味着 cd my-project && rbrain stats 无需任何参数即可使用项目知识库。
Skill / Claude Code 技能
研究工作流技能文件位于 rbrain-research-cli/SKILL.md,涵盖完整工作流:知识库发现、检索、写入、图谱链接、Dream Cycle、Signal Detection 和 MCP 使用模式。
A Claude Code skill for research workflows is at rbrain-research-cli/SKILL.md. Covers brain discovery, retrieval, writing, graph linking, dream cycle, signal detection, and MCP usage patterns.