rag - SKILL.md Agent Skill

name: rag description: > Manage RAG (Retrieval-Augmented Generation) knowledge sources for agents. Add external documentation (GitHub repos, doc sites) to an agent's local knowledge base so memory_search can retrieve domain-specific information. Triggers when: user says agent's knowledge is inaccurate, user wants agent to learn about a topic, user asks to add/remove/list RAG sources, user mentions "RAG", "知识库", "加个文档", "学习一下", "知识不准", "回答不对", "补充知识".

RAG Knowledge Source Management

Manage local RAG knowledge sources for OpenClaw agents. Sources are stored in each agent's <workspace>/rag/ directory and automatically indexed by OpenClaw's memorySearch via extraPaths: ["rag"].

How It Works

Each agent has a rag/ folder in its workspace
OpenClaw's memorySearch indexes all .md files in this folder
Sources are typically git repos cloned with --depth 1 (shallow)
A single cron job per agent periodically git pull all sources
After adding/removing sources, a gateway restart triggers reindexing

Add a RAG Source

Step 1: Confirm with User

Before doing anything, send a briefing to the user:

📚 RAG 知识库建立方案

要给 [agent名称] 添加 [topic] 的知识库，具体操作如下：

1️⃣ 下载文档
   将 [repo/source] 克隆到 <workspace>/rag/ 目录下
   预计包含约 X 个 Markdown 文件

2️⃣ 建立定时更新
   [如果该 agent 还没有 rag-auto-update 定时任务]
   创建每日自动更新任务，定期拉取最新文档
   [如果已有 rag-auto-update 任务]
   将新源合并到现有的 rag-auto-update 定时任务中

3️⃣ 重启生效
   需要重启 gateway，等待索引完成即可

是否继续？

等用户确认后再执行。不要跳过这一步。

Step 2: Clone the Source

mkdir -p <workspace>/rag
cd <workspace>/rag
git clone --depth 1 <repo-url> [folder-name]

Verify useful .md files exist:

find <workspace>/rag/<folder> -name "*.md" | wc -l

Step 3: Set Up or Update Cron Job

关键规则：每个 agent 只有一个 rag-auto-update 定时任务。

Check if the agent already has a rag-auto-update cron job:

cat ~/.openclaw/cron/jobs.json | python3 -c "
import json,sys
d=json.load(sys.stdin)
for j in d['jobs']:
    if j.get('agentId')=='<AGENT_ID>' and j['name']=='rag-auto-update':
        print('FOUND:', j['id'])
        print(j['payload']['message'])
"

If no existing job: Create one that covers all sources in <workspace>/rag/:

The cron job message should:

List all rag/ subdirectories
Run git pull on each one
Report which ones had updates

Example cron payload message:

执行 RAG 知识源更新。

1. 获取所有 RAG 源目录：ls <workspace>/rag/
2. 对每个目录执行 git pull
3. 收集更新结果：
   - 如果全部 "Already up to date." → 回复 NO_REPLY
   - 如果有更新 → 汇总变更，用 message 发简报给用户（target: <USER_ID>）：
     📦 RAG 知识源更新
     📂 [source1]: 更新了 X 个文件
     📂 [source2]: 无变化
     发完回复 NO_REPLY
4. 记录到 memory/YYYY-MM-DD.md

Cron job settings:

name: rag-auto-update
schedule: 0 3 * * * (每天凌晨 3:00)
sessionTarget: isolated
wakeMode: now
delivery.mode: announce
delivery.channel: telegram
delivery.to: user's Telegram ID

If job already exists: Update its payload message to scan all directories under rag/ dynamically. Since the job already lists and iterates all rag/ subdirectories, it will automatically pick up new sources. Only update the job if it hardcodes specific folder names instead of scanning dynamically.

Step 4: Notify User

Tell the user:

✅ 文档已下载到 <workspace>/rag/<folder>/
✅ 定时更新已配置（每天凌晨 3:00 自动拉取）
⚠️ 请重启 gateway，等待索引完成即可

List RAG Sources

ls <workspace>/rag/

For each source, show:

Folder name
Git remote URL: git -C <path> remote get-url origin
Last update: git -C <path> log -1 --format="%ai"
File count: find <path> -name "*.md" | wc -l

Remove a RAG Source

Confirm with user before removing
Remove: rm -rf <workspace>/rag/<source-name>
No need to update cron job (it scans dynamically)
Tell user to restart gateway to update the index

Update a RAG Source (Manual)

cd <workspace>/rag/<source-name> && git pull

The memorySearch file watcher detects changes and reindexes automatically.

Notes

Only .md files are indexed by memorySearch
Git repos use --depth 1 to save space
rag/ directories are gitignored per workspace (workspaces/*/rag/)
Each agent's index is separate (~/.openclaw/memory/<agentId>.sqlite)
For non-git sources, manually download markdown files into rag/