| name | knowledge-query |
| description | Query CA knowledge graph before answering knowledge questions. ONLY for concept explanations. MUST extract core topic (2-10 words), NOT entire sentences. MUST NOT use for commands, code, debugging, tasks. |
knowledge-query
Query CA (Curious Agent) knowledge graph to check confidence level before answering knowledge questions.
⚠️ CRITICAL: Topic Extraction Rules
YOU MUST extract the CORE topic, NOT the full user sentence.
| User Message | WRONG Topic ❌ | RIGHT Topic ✅ |
|---|
| "什么是agent上下文管理系统？" | "什么是agent上下文管理系统？" | "agent上下文管理系统" |
| "解释一下transformer attention机制" | "解释一下transformer attention机制" | "transformer attention" |
| "agent记忆机制是怎么实现的？" | "agent记忆机制是怎么实现的？" | "agent记忆机制" |
| "帮我看看RAG和vector search的区别" | "帮我看看RAG和vector search的区别" | "RAG" 或 "vector search" |
Topic Format Requirements:
- Length: 2-10 words (noun phrases, not questions)
- Language: Chinese or English keywords only
- No question words: 不要包含"什么是"、"解释"、"怎么"、"帮我"等
- No command words: 不要包含"检查"、"commit"、"push"、"debug"等
✅ WHEN TO USE (Must Invoke)
Invoke this skill ONLY when user asks about:
- 概念解释: "什么是 X", "X 是什么", "解释 X", "X 的原理"
- 技术对比: "X 和 Y 的区别", "比较 X 和 Y"
- 机制说明: "X 是怎么工作的", "X 的实现原理"
- 名词询问: 用户提到不熟悉的技术术语
Example valid triggers:
- "什么是 FlashAttention？"
- "agent memory 有哪些类型？"
- "RAG 和 semantic search 有什么区别？"
❌ MUST NOT USE (Skip This Skill)
NEVER invoke this skill for:
| Category | Examples | Reason |
|---|
| 命令执行 | "ls", "git status", "npm install", "重启服务" | These are operations, not knowledge |
| 代码编写 | "写一个函数", "帮我写代码", "实现 X 功能" | Code generation, not explanation |
| 问题调试 | "为什么报错", "debug 这个问题", "检查一下" | Troubleshooting, not knowledge |
| 任务执行 | "commit", "push", "创建 PR", "部署" | Task execution, not explanation |
| 日常对话 | "好的", "谢谢", "继续" | No knowledge query needed |
| 文件操作 | "读一下这个文件", "修改配置" | File operations, not knowledge |
If user's intent is ANY of the above → SKIP this skill entirely.
How to Use
Step 1: Extract Core Topic
Before calling the script, extract the core topic from user's message:
Step 2: Run Query Script
python3 /root/.openclaw/skills/knowledge-query/scripts/query.py "<extracted_topic>"
Script path alternatives (if not installed):
- CA project:
/root/dev/curious-agent/openclaw-skills/knowledge-query/scripts/query.py
Step 3: Process Output
Script returns JSON:
{
"success": true,
"output": "[KG Context — Expert (85%)]\n话题: agent上下文管理系统\n置信度: 0.85\nKG 有完整知识，可直接回答。",
"metadata": {"topic": "...", "confidence": 0.85, "level": "expert", "gaps": []}
}
Read the output field and follow guidance:
| Level | Confidence | Your Action |
|---|
| 🟢 Expert | ≥85% | Answer from KG knowledge directly, cite sources |
| 🟡 Intermediate | 60-85% | Use KG knowledge + supplement with web search |
| 🟠 Beginner | 30-60% | Search first, then answer with combined knowledge |
| 🔴 Novice | <30% | Answer from LLM, optionally inject topic to CA for exploration |
Execution Checklist
Before invoking, verify ALL conditions:
[ ] User asks about a concept/explanation/mechanism (NOT command/code/debug)
[ ] Extracted topic is 2-10 words (noun phrase, no question words)
[ ] Topic is NOT a command keyword (git, commit, ls, debug, etc.)
If ANY condition fails → SKIP this skill and proceed directly.
Requirements
- CA API running on
localhost:4848 (env CA_API_URL overrides)
- Script handles errors gracefully (returns warning if API down)
- Timeout: 2 seconds (non-blocking)
Error Handling
If success: false:
- Log:
[KG Context — 不可用] CA API 无响应，跳过知识查询。
- Proceed with normal answer (do NOT retry, do NOT block)
