vibeCodingP2 / experiments /skill_knowledge_router.py
dellweb3's picture
Add Knowledge routing and semantic search experiments
2f6097e verified
Raw History Blame Contribute Delete
9.47 kB
"""
ZKCoding V0.1
Skill + Knowledge 智能推荐原型
运行:
python experiments/skill_knowledge_router.py
运行自测:
python experiments/skill_knowledge_router.py --self-test
"""
import sys
# 1. Demo 中现有的 Skill
SKILLS = {
"general": "普通对话(不执行工具)",
"erp": "ERP 订单助手",
"wms": "WMS 库存助手",
"device": "设备工单助手",
"java_erp": "Java 编程 · ERP 采购报价",
"java_wms": "Java 编程 · WMS 库存流水",
"java_device": "Java 编程 · 设备工单与巡检",
"java_calculator": "Java 编程 · 网页计算器",
"java_pdf": "Java 编程 · PDF读取与解析",
"verified_erp": "Java 编程 · ERP 采购报价 · 已验证源码技能",
}
# 2. Demo 中现有的 Knowledge
KNOWLEDGE = {
"rules": "Java 编码规约",
"basics": "Java 基础知识",
"integration": "Java 集成知识",
"erp": "ERP 订单知识库",
"wms": "WMS 库存知识库",
"device": "设备工单知识库",
"sap": "SAP 接口参考",
}
# 3. 按业务领域整理特征词
# 分数越高,说明该词越有辨识度
DOMAIN_WORDS = {
"erp": {
"erp": 4,
"采购报价": 5,
"采购订单": 5,
"供应商": 4,
"采购": 3,
"报价": 3,
"审批": 2,
"收货": 2,
"付款": 2,
"订单": 1,
},
"wms": {
"wms": 4,
"库存流水": 5,
"库存": 4,
"仓库": 3,
"入库": 4,
"出库": 4,
"盘点": 3,
"库存余额": 5,
"商品进出": 3,
"剩余数量": 2,
"sku": 4,
},
"device": {
"设备运维": 5,
"设备巡检": 5,
"设备工单": 5,
"巡检": 4,
"维修": 3,
"工单": 3,
"运维": 3,
"故障": 2,
"设备": 2,
},
"calculator": {
"计算器": 5,
"网页计算器": 6,
"四则运算": 4,
},
"pdf": {
"pdf": 5,
"pdfbox": 5,
"文档解析": 4,
"文件解析": 3,
"提取pdf": 5,
},
}
# 4. 判断用户是否需要生成代码
CODING_WORDS = [
"帮我写", "写一个", "生成代码",
"编写代码", "开发一个", "实现一个",
"帮我开发", "创建一个", "搭建一个",
"修改代码", "生成java", "写java",
"java项目", "java程序", "java网站",
"spring boot项目", "编程实现",
"开发", "编写", "生成代码",
"实现", "写代码", "构建",
]
# Java 技术问题,但不一定需要执行编程 Skill
JAVA_WORDS = [
"java", "spring", "spring boot", "maven",
"jdk", "junit", "jdbc", "controller",
"rest api", "接口", "类名", "方法名",
]
INTEGRATION_WORDS = [
"api", "rest", "http", "json", "maven",
"依赖", "文件解析", "pdfbox",
"集成", "接口对接", "jdbc",
]
def get_domain_scores(question):
"""给每个业务领域打分。"""
text = question.lower()
scores = {}
for domain, words in DOMAIN_WORDS.items():
score = 0
for word, weight in words.items():
if word.lower() in text:
score += weight
scores[domain] = score
return scores
def recommend(question):
"""根据问题推荐 Skill 和 Knowledge。"""
text = question.strip().lower()
if not text:
return {
"skill": None,
"knowledge": [],
"domain": None,
"confidence": "低",
"reasons": ["请先输入需求"],
"needs_confirmation": True,
}
scores = get_domain_scores(text)
# 找到最可能的业务领域
ranked = sorted(
scores,
key=lambda name: scores[name],
reverse=True,
)
domain = ranked[0] if scores[ranked[0]] > 0 else None
top_score = scores[domain] if domain else 0
is_coding = any(
word in text for word in CODING_WORDS
)
is_java = any(
word in text for word in JAVA_WORDS
)
is_integration = any(
word in text for word in INTEGRATION_WORDS
)
reasons = []
# 5. 推荐 Skill
skill = SKILLS["general"]
if is_coding:
coding_map = {
"erp": "java_erp",
"wms": "java_wms",
"device": "java_device",
"calculator": "java_calculator",
"pdf": "java_pdf",
}
if domain in coding_map:
skill = SKILLS[coding_map[domain]]
reasons.append(
"识别为编程任务,推荐对应 Java 编程 Skill"
)
else:
reasons.append(
"识别到编程需求,但没有足够信息匹配具体 Skill"
)
elif domain in ("erp", "wms", "device"):
skill = SKILLS[domain]
reasons.append(
"识别为业务查询或业务操作,推荐业务助手"
)
else:
reasons.append(
"暂不需要执行特定 Skill,使用普通对话"
)
# 6. 推荐 Knowledge
knowledge = []
def add_knowledge(key):
name = KNOWLEDGE[key]
if name not in knowledge:
knowledge.append(name)
if is_coding:
add_knowledge("rules")
reasons.append("编程任务需要参考 Java 编码规约")
if is_java:
add_knowledge("basics")
if is_integration:
add_knowledge("integration")
# 同时涉及多个业务领域时,允许推荐多个知识库
for name in ("erp", "wms", "device"):
if scores[name] > 0:
add_knowledge(name)
if "sap" in text:
add_knowledge("sap")
reasons.append("检测到 SAP 接口相关需求")
# 7. 可复用的已验证源码技能
if (
domain == "erp"
and is_coding
and any(
word in text
for word in ["复用", "已验证", "已有源码"]
)
):
skill = SKILLS["verified_erp"]
reasons.append(
"用户希望复用已验证的 ERP 源码技能"
)
# 8. 判断推荐是否需要进一步确认
tied = (
domain is not None
and sum(
1 for score in scores.values()
if score == top_score
) > 1
)
if domain is None:
confidence = "低"
elif tied or top_score < 3:
confidence = "低"
elif top_score < 6:
confidence = "中"
else:
confidence = "高"
needs_confirmation = (
confidence != "高"
or (is_coding and domain is None)
)
# 目前置信度仅基于规则分数,不是统计概率
return {
"skill": skill,
"knowledge": knowledge,
"domain": domain,
"confidence": confidence,
"reasons": reasons,
"needs_confirmation": needs_confirmation,
}
def show_result(result):
"""展示推荐结果。"""
print("\n========== ZKCoding 推荐结果 ==========")
print("Skill:", result["skill"] or "暂无")
print(
"Knowledge:",
"、".join(result["knowledge"]) or "暂无推荐",
)
print("匹配领域:", result["domain"] or "未识别")
print("规则匹配强度:", result["confidence"])
print("\n推荐原因:")
for reason in result["reasons"]:
print("-", reason)
print("\n注意:建议用户确认后再执行。")
print("=======================================\n")
def run_tests():
"""简单自测,不调用任何模型。"""
cases = [
(
"帮我用 Java 开发 ERP 采购报价系统",
"java_erp",
"erp",
),
(
"帮我写一个 WMS 库存流水接口",
"java_wms",
"wms",
),
(
"帮我开发设备巡检工单系统",
"java_device",
"device",
),
(
"帮我开发一个网页计算器",
"java_calculator",
"calculator",
),
(
"帮我开发一个 PDF 解析程序",
"java_pdf",
"pdf",
),
(
"查询采购订单的付款状态",
"erp",
"erp",
),
(
"查询仓库库存余额",
"wms",
"wms",
),
(
"查询设备维修工单",
"device",
"device",
),
]
success = 0
for question, expected_skill, expected_domain in cases:
result = recommend(question)
expected_name = SKILLS[expected_skill]
correct = (
result["skill"] == expected_name
and result["domain"] == expected_domain
)
print("问题:", question)
print("结果:", "PASS" if correct else "FAIL")
print("推荐:", result["skill"], "\n")
if correct:
success += 1
print(f"自测结果: {success}/{len(cases)}")
return success == len(cases)
if __name__ == "__main__":
if "--self-test" in sys.argv:
if not run_tests():
sys.exit(1)
else:
while True:
question = input(
"请输入需求(输入 exit 退出):"
).strip()
if question.lower() == "exit":
break
result = recommend(question)
show_result(result)