dellweb3 commited on
Commit
2f6097e
·
verified ·
1 Parent(s): d437033

Add Knowledge routing and semantic search experiments

Browse files
experiments/hybrid_search.py ADDED
@@ -0,0 +1,166 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ import re
2
+ from sentence_transformers import SentenceTransformer, util
3
+
4
+ # 1. 使用已经下载过的语义模型
5
+ MODEL_NAME = "intfloat/multilingual-e5-small"
6
+
7
+ # 2. 模拟企业知识库(每份文档都有 ID)
8
+ documents = [
9
+ {
10
+ "id": "WMS-001",
11
+ "text": "库存管理系统记录商品入库、出库和剩余库存数量。",
12
+ },
13
+ {
14
+ "id": "ERP-001",
15
+ "text": "ERP 采购订单系统负责供应商报价、采购审批和付款流程。",
16
+ },
17
+ {
18
+ "id": "JAVA-001",
19
+ "text": "Java Spring Boot 通过 REST Controller 实现 HTTP API。",
20
+ },
21
+ {
22
+ "id": "DEVICE-001",
23
+ "text": "设备运维系统负责设备巡检、维修工单与故障记录。",
24
+ },
25
+ {
26
+ "id": "JAVA-002",
27
+ "text": "Java 中的 inventoryTransaction 用于库存事务处理。",
28
+ },
29
+ ]
30
+
31
+ # 3. 提取检索词:中文双字组合 + 英文单词
32
+ def tokenize(text):
33
+ text = text.lower()
34
+ tokens = set(re.findall(r"[a-z][a-z0-9_]*", text))
35
+
36
+ for segment in re.findall(r"[\u4e00-\u9fff]+", text):
37
+ for i in range(len(segment) - 1):
38
+ tokens.add(segment[i:i + 2])
39
+
40
+ return tokens
41
+
42
+
43
+ # 4. 关键词检索(简化实验版)
44
+ def keyword_search(query):
45
+ query_tokens = tokenize(query)
46
+ results = []
47
+
48
+ for doc in documents:
49
+ doc_tokens = tokenize(doc["text"])
50
+ score = len(query_tokens & doc_tokens)
51
+
52
+ if score > 0:
53
+ results.append((doc["id"], score))
54
+
55
+ return sorted(
56
+ results,
57
+ key=lambda item: item[1],
58
+ reverse=True,
59
+ )
60
+
61
+
62
+ # 5. 加载语义模型
63
+ print("正在加载 Embedding 模型...")
64
+ model = SentenceTransformer(MODEL_NAME)
65
+
66
+ passages = [
67
+ "passage: " + doc["text"]
68
+ for doc in documents
69
+ ]
70
+
71
+ doc_vectors = model.encode(
72
+ passages,
73
+ convert_to_tensor=True,
74
+ normalize_embeddings=True,
75
+ )
76
+
77
+
78
+ # 6. 语义检索
79
+ def semantic_search(query):
80
+ query_vector = model.encode(
81
+ "query: " + query,
82
+ convert_to_tensor=True,
83
+ normalize_embeddings=True,
84
+ )
85
+
86
+ similarities = util.cos_sim(
87
+ query_vector, doc_vectors
88
+ )[0]
89
+
90
+ results = [
91
+ (doc["id"], float(similarities[i]))
92
+ for i, doc in enumerate(documents)
93
+ ]
94
+
95
+ return sorted(
96
+ results,
97
+ key=lambda item: item[1],
98
+ reverse=True,
99
+ )
100
+
101
+
102
+ # 7. 混合检索:按排名融合
103
+ def hybrid_search(query):
104
+ keyword_results = keyword_search(query)
105
+ semantic_results = semantic_search(query)
106
+
107
+ # RRF:排名越靠前,得分越高
108
+ scores = {}
109
+ k = 60
110
+
111
+ for rank, (doc_id, _) in enumerate(
112
+ keyword_results, start=1
113
+ ):
114
+ scores[doc_id] = (
115
+ scores.get(doc_id, 0) + 1 / (k + rank)
116
+ )
117
+
118
+ for rank, (doc_id, _) in enumerate(
119
+ semantic_results, start=1
120
+ ):
121
+ scores[doc_id] = (
122
+ scores.get(doc_id, 0) + 1 / (k + rank)
123
+ )
124
+
125
+ results = sorted(
126
+ scores.items(),
127
+ key=lambda item: item[1],
128
+ reverse=True,
129
+ )
130
+
131
+ return keyword_results, semantic_results, results
132
+
133
+
134
+ # 8. 展示搜索结果
135
+ def display(query):
136
+ kw, sem, hybrid = hybrid_search(query)
137
+
138
+ print("\n" + "=" * 55)
139
+ print("用户问题:", query)
140
+
141
+ print("\n【关键词检索 Top 3】")
142
+ for doc_id, score in kw[:3]:
143
+ print(doc_id, "词项得分:", score)
144
+
145
+ print("\n【语义检索 Top 3】")
146
+ for doc_id, score in sem[:3]:
147
+ print(doc_id, "相似度:", round(score, 4))
148
+
149
+ print("\n【混合检索 Top 3】")
150
+ for doc_id, score in hybrid[:3]:
151
+ doc = next(
152
+ d for d in documents if d["id"] == doc_id
153
+ )
154
+ print(doc_id, "RRF:", round(score, 5))
155
+ print(" 文档:", doc["text"])
156
+
157
+
158
+ # 9. 测试三个场景
159
+ questions = [
160
+ "如何追踪商品的进出及剩余数量?",
161
+ "怎样管理供应商报价和采购审批?",
162
+ "Spring Boot 如何编写 REST API?",
163
+ ]
164
+
165
+ for question in questions:
166
+ display(question)
experiments/knowledge_integration_test.py ADDED
@@ -0,0 +1,106 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+
2
+ """
3
+ ZKCoding Knowledge + Embedding 集成实验
4
+
5
+ 使用原项目的 KnowledgeCatalog 和 SQLite Repository,
6
+ 然后调用自己编写的 KnowledgeReranker。
7
+
8
+ 运行:
9
+ python -m experiments.knowledge_integration_test
10
+ """
11
+
12
+ import tempfile
13
+ from pathlib import Path
14
+
15
+ from src.adapters.indexes.knowledge_sqlite import (
16
+ SqliteKnowledgeRepository,
17
+ )
18
+ from src.modules.knowledge.application.catalog import (
19
+ KnowledgeCatalog,
20
+ )
21
+ from src.modules.skill.domain.models import Actor
22
+
23
+ from experiments.knowledge_reranker import KnowledgeReranker
24
+
25
+
26
+ def main():
27
+ # 1. 创建测试用户
28
+ actor = Actor(
29
+ tenant_id="test-tenant",
30
+ user_id="developer",
31
+ roles=[],
32
+ )
33
+
34
+ # 2. 创建临时数据库
35
+ with tempfile.TemporaryDirectory() as temp_dir:
36
+ folder = Path(temp_dir)
37
+
38
+ repo = SqliteKnowledgeRepository(
39
+ folder / "index.sqlite"
40
+ )
41
+
42
+ # 3. 复用组长已有的测试数据生成函数
43
+ from src.modules.knowledge.tests.test_knowledge import (
44
+ write_fixture,
45
+ )
46
+
47
+ corpus_dir = folder / "corpus"
48
+ catalog_path = write_fixture(corpus_dir)
49
+
50
+ repo.import_corpus(
51
+ corpus_dir,
52
+ actor.tenant_id,
53
+ catalog_path,
54
+ )
55
+
56
+ # 4. 创建真正的 Knowledge Service
57
+ catalog = KnowledgeCatalog(repo)
58
+
59
+ # 5. 调用原有 SQLite FTS 检索
60
+ query = "inventory"
61
+
62
+ records = catalog.search(
63
+ actor,
64
+ query,
65
+ ["fixture-java"],
66
+ )
67
+
68
+ print("\n【SQLite 检索结果】")
69
+ print("找到文档数量:", len(records))
70
+
71
+ for record in records:
72
+ print(
73
+ record["citation_id"],
74
+ record["title"],
75
+ record["source_url"],
76
+ )
77
+
78
+ # 6. 调用我们新增的 Embedding 重排序
79
+ reranker = KnowledgeReranker()
80
+
81
+ ranked = reranker.rerank(
82
+ query,
83
+ records,
84
+ )
85
+
86
+ print("\n【Embedding 重排序结果】")
87
+
88
+ for record in ranked:
89
+ print(
90
+ record["citation_id"],
91
+ "相似度:",
92
+ round(record["semantic_score"], 4),
93
+ "来源:",
94
+ record["source_url"],
95
+ )
96
+
97
+ # 7. 验证引用信息没有丢失
98
+ assert len(ranked) == len(records)
99
+ assert all(
100
+ r["source_url"] for r in ranked
101
+ )
102
+ print("\n集成测试通过:搜索与重排序已连接!")
103
+
104
+
105
+ if __name__ == "__main__":
106
+ main()
experiments/knowledge_reranker.py ADDED
@@ -0,0 +1,83 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ from sentence_transformers import SentenceTransformer
2
+
3
+ MODEL_NAME = "intfloat/multilingual-e5-small"
4
+
5
+
6
+ class KnowledgeReranker:
7
+ def __init__(self):
8
+ self.model = SentenceTransformer(MODEL_NAME)
9
+
10
+ def rerank(self, query, records):
11
+ # 没有找到文档时,直接返回空列表
12
+ if not records:
13
+ return []
14
+
15
+ # 提取现有 Knowledge Service 返回的内容
16
+ passages = [
17
+ "passage: " + r["title"] + " " + r["snippet"]
18
+ for r in records
19
+ ]
20
+
21
+ # 将问题和文档转换为向量
22
+ query_vector = self.model.encode(
23
+ "query: " + query,
24
+ normalize_embeddings=True,
25
+ )
26
+
27
+ document_vectors = self.model.encode(
28
+ passages,
29
+ normalize_embeddings=True,
30
+ )
31
+
32
+ # 计算语义相似度
33
+ scores = document_vectors @ query_vector
34
+
35
+ # 保留原始来源、版本、许可证等字段
36
+ ranked = sorted(
37
+ zip(records, scores),
38
+ key=lambda item: float(item[1]),
39
+ reverse=True,
40
+ )
41
+
42
+ return [
43
+ {**record, "semantic_score": float(score)}
44
+ for record, score in ranked
45
+ ]
46
+
47
+
48
+ if __name__ == "__main__":
49
+ # 模拟现有 Knowledge Service 已经授权返回的结果
50
+ records = [
51
+ {
52
+ "citation_id": "WMS-001",
53
+ "title": "库存管理",
54
+ "snippet": "商品入库、出库和库存余额管理。",
55
+ "source_url": "example://wms",
56
+ },
57
+ {
58
+ "citation_id": "ERP-001",
59
+ "title": "采购管理",
60
+ "snippet": "供应商报价、订单审批和付款管理。",
61
+ "source_url": "example://erp",
62
+ },
63
+ {
64
+ "citation_id": "JAVA-001",
65
+ "title": "Java API",
66
+ "snippet": "Spring Boot REST API 接口开发。",
67
+ "source_url": "example://java",
68
+ },
69
+ ]
70
+
71
+ reranker = KnowledgeReranker()
72
+
73
+ result = reranker.rerank(
74
+ "怎样管理供应商采购报价?",
75
+ records,
76
+ )
77
+
78
+ for item in result:
79
+ print(
80
+ item["citation_id"],
81
+ round(item["semantic_score"], 4),
82
+ item["source_url"],
83
+ )
experiments/semantic_search.py ADDED
@@ -0,0 +1,47 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ from sentence_transformers import SentenceTransformer, util
2
+
3
+ # 加载支持中英文的 Embedding 模型
4
+ MODEL_NAME = "intfloat/multilingual-e5-small"
5
+
6
+ print("正在加载语义模型(第一次运行需要下载)...")
7
+ model = SentenceTransformer(MODEL_NAME)
8
+
9
+ # 模拟 ZKCoding 的 Java / ERP / WMS 知识库
10
+ documents = [
11
+ "库存管理系统记录商品入库、出库和剩余库存数量。",
12
+ "ERP 采购订单系统负责供应商报价、采购审批和付款流程。",
13
+ "Java Spring Boot 可以通过 REST Controller 实现 HTTP API。",
14
+ "设备运维系统负责设备巡检、维修工单及故障记录。",
15
+ ]
16
+
17
+ # E5 模型使用 passage: 和 query: 前缀
18
+ passages = ["passage: " + doc for doc in documents]
19
+ document_embeddings = model.encode(
20
+ passages,
21
+ convert_to_tensor=True,
22
+ normalize_embeddings=True,
23
+ )
24
+
25
+ questions = [
26
+ "如何追踪商品的进出及剩余数量?",
27
+ "怎样管理供应商的报价和采购流程?",
28
+ "Spring Boot 如何编写接口?",
29
+ ]
30
+
31
+ for question in questions:
32
+ query_embedding = model.encode(
33
+ "query: " + question,
34
+ convert_to_tensor=True,
35
+ normalize_embeddings=True,
36
+ )
37
+
38
+ scores = util.cos_sim(
39
+ query_embedding,
40
+ document_embeddings,
41
+ )[0]
42
+
43
+ best_index = int(scores.argmax())
44
+
45
+ print("\n问题:", question)
46
+ print("最相关文档:", documents[best_index])
47
+ print("相似度:", round(float(scores[best_index]), 4))
experiments/skill_knowledge_router.py ADDED
@@ -0,0 +1,390 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ """
2
+ ZKCoding V0.1
3
+ Skill + Knowledge 智能推荐原型
4
+
5
+ 运行:
6
+ python experiments/skill_knowledge_router.py
7
+
8
+ 运行自测:
9
+ python experiments/skill_knowledge_router.py --self-test
10
+
11
+ """
12
+
13
+ import sys
14
+
15
+ # 1. Demo 中现有的 Skill
16
+ SKILLS = {
17
+ "general": "普通对话(不执行工具)",
18
+ "erp": "ERP 订单助手",
19
+ "wms": "WMS 库存助手",
20
+ "device": "设备工单助手",
21
+ "java_erp": "Java 编程 · ERP 采购报价",
22
+ "java_wms": "Java 编程 · WMS 库存流水",
23
+ "java_device": "Java 编程 · 设备工单与巡检",
24
+ "java_calculator": "Java 编程 · 网页计算器",
25
+ "java_pdf": "Java 编程 · PDF读取与解析",
26
+ "verified_erp": "Java 编程 · ERP 采购报价 · 已验证源码技能",
27
+ }
28
+
29
+
30
+ # 2. Demo 中现有的 Knowledge
31
+ KNOWLEDGE = {
32
+ "rules": "Java 编码规约",
33
+ "basics": "Java 基础知识",
34
+ "integration": "Java 集成知识",
35
+ "erp": "ERP 订单知识库",
36
+ "wms": "WMS 库存知识库",
37
+ "device": "设备工单知识库",
38
+ "sap": "SAP 接口参考",
39
+ }
40
+
41
+
42
+ # 3. 按业务领域整理特征词
43
+ # 分数越高,说明该词越有辨识度
44
+ DOMAIN_WORDS = {
45
+ "erp": {
46
+ "erp": 4,
47
+ "采购报价": 5,
48
+ "采购订单": 5,
49
+ "供应商": 4,
50
+ "采购": 3,
51
+ "报价": 3,
52
+ "审批": 2,
53
+ "收货": 2,
54
+ "付款": 2,
55
+ "订单": 1,
56
+ },
57
+ "wms": {
58
+ "wms": 4,
59
+ "库存流水": 5,
60
+ "库存": 4,
61
+ "仓库": 3,
62
+ "入库": 4,
63
+ "出库": 4,
64
+ "盘点": 3,
65
+ "库存余额": 5,
66
+ "商品进出": 3,
67
+ "剩余数量": 2,
68
+ "sku": 4,
69
+ },
70
+ "device": {
71
+ "设备运维": 5,
72
+ "设备巡检": 5,
73
+ "设备工单": 5,
74
+ "巡检": 4,
75
+ "维修": 3,
76
+ "工单": 3,
77
+ "运维": 3,
78
+ "故障": 2,
79
+ "设备": 2,
80
+ },
81
+ "calculator": {
82
+ "计算器": 5,
83
+ "网页计算器": 6,
84
+ "四则运算": 4,
85
+ },
86
+ "pdf": {
87
+ "pdf": 5,
88
+ "pdfbox": 5,
89
+ "文档解析": 4,
90
+ "文件解析": 3,
91
+ "提取pdf": 5,
92
+ },
93
+ }
94
+
95
+
96
+ # 4. 判断用户是否需要生成代码
97
+ CODING_WORDS = [
98
+ "帮我写", "写一个", "生成代码",
99
+ "编写代码", "开发一个", "实现一个",
100
+ "帮我开发", "创建一个", "搭建一个",
101
+ "修改代码", "生成java", "写java",
102
+ "java项目", "java程序", "java网站",
103
+ "spring boot项目", "编程实现",
104
+ "开发", "编写", "生成代码",
105
+ "实现", "写代码", "构建",
106
+ ]
107
+
108
+ # Java 技术问题,但不一定需要执行编程 Skill
109
+ JAVA_WORDS = [
110
+ "java", "spring", "spring boot", "maven",
111
+ "jdk", "junit", "jdbc", "controller",
112
+ "rest api", "接口", "类名", "方法名",
113
+ ]
114
+
115
+ INTEGRATION_WORDS = [
116
+ "api", "rest", "http", "json", "maven",
117
+ "依赖", "文件解析", "pdfbox",
118
+ "集成", "接口对接", "jdbc",
119
+ ]
120
+
121
+
122
+ def get_domain_scores(question):
123
+ """给每个业务领域打分。"""
124
+ text = question.lower()
125
+ scores = {}
126
+
127
+ for domain, words in DOMAIN_WORDS.items():
128
+ score = 0
129
+
130
+ for word, weight in words.items():
131
+ if word.lower() in text:
132
+ score += weight
133
+
134
+ scores[domain] = score
135
+
136
+ return scores
137
+
138
+
139
+ def recommend(question):
140
+ """根据问题推荐 Skill 和 Knowledge。"""
141
+
142
+ text = question.strip().lower()
143
+
144
+ if not text:
145
+ return {
146
+ "skill": None,
147
+ "knowledge": [],
148
+ "domain": None,
149
+ "confidence": "低",
150
+ "reasons": ["请先输入需求"],
151
+ "needs_confirmation": True,
152
+ }
153
+
154
+ scores = get_domain_scores(text)
155
+
156
+ # 找到最可能的业务领域
157
+ ranked = sorted(
158
+ scores,
159
+ key=lambda name: scores[name],
160
+ reverse=True,
161
+ )
162
+
163
+ domain = ranked[0] if scores[ranked[0]] > 0 else None
164
+ top_score = scores[domain] if domain else 0
165
+
166
+ is_coding = any(
167
+ word in text for word in CODING_WORDS
168
+ )
169
+
170
+ is_java = any(
171
+ word in text for word in JAVA_WORDS
172
+ )
173
+
174
+ is_integration = any(
175
+ word in text for word in INTEGRATION_WORDS
176
+ )
177
+
178
+ reasons = []
179
+
180
+ # 5. 推荐 Skill
181
+ skill = SKILLS["general"]
182
+
183
+ if is_coding:
184
+ coding_map = {
185
+ "erp": "java_erp",
186
+ "wms": "java_wms",
187
+ "device": "java_device",
188
+ "calculator": "java_calculator",
189
+ "pdf": "java_pdf",
190
+ }
191
+
192
+ if domain in coding_map:
193
+ skill = SKILLS[coding_map[domain]]
194
+ reasons.append(
195
+ "识别为编程任务,推荐对应 Java 编程 Skill"
196
+ )
197
+ else:
198
+ reasons.append(
199
+ "识别到编程需求,但没有足够信息匹配具体 Skill"
200
+ )
201
+
202
+ elif domain in ("erp", "wms", "device"):
203
+ skill = SKILLS[domain]
204
+ reasons.append(
205
+ "识别为业务查询或业务操作,推荐业务助手"
206
+ )
207
+
208
+ else:
209
+ reasons.append(
210
+ "暂不需要执行特定 Skill,使用普通对话"
211
+ )
212
+
213
+ # 6. 推荐 Knowledge
214
+ knowledge = []
215
+
216
+ def add_knowledge(key):
217
+ name = KNOWLEDGE[key]
218
+ if name not in knowledge:
219
+ knowledge.append(name)
220
+
221
+ if is_coding:
222
+ add_knowledge("rules")
223
+ reasons.append("编程任务需要参考 Java 编码规约")
224
+
225
+ if is_java:
226
+ add_knowledge("basics")
227
+
228
+ if is_integration:
229
+ add_knowledge("integration")
230
+
231
+ # 同时涉及多个业务领域时,允许推荐多个知识库
232
+ for name in ("erp", "wms", "device"):
233
+ if scores[name] > 0:
234
+ add_knowledge(name)
235
+
236
+ if "sap" in text:
237
+ add_knowledge("sap")
238
+ reasons.append("检测到 SAP 接口相关需求")
239
+
240
+ # 7. 可复用的已验证源码技能
241
+ if (
242
+ domain == "erp"
243
+ and is_coding
244
+ and any(
245
+ word in text
246
+ for word in ["复用", "已验证", "已有源码"]
247
+ )
248
+ ):
249
+ skill = SKILLS["verified_erp"]
250
+ reasons.append(
251
+ "用户希望复用已验证的 ERP 源码技能"
252
+ )
253
+
254
+ # 8. 判断推荐是否需要进一步确认
255
+ tied = (
256
+ domain is not None
257
+ and sum(
258
+ 1 for score in scores.values()
259
+ if score == top_score
260
+ ) > 1
261
+ )
262
+
263
+ if domain is None:
264
+ confidence = "低"
265
+ elif tied or top_score < 3:
266
+ confidence = "低"
267
+ elif top_score < 6:
268
+ confidence = "中"
269
+ else:
270
+ confidence = "高"
271
+
272
+ needs_confirmation = (
273
+ confidence != "高"
274
+ or (is_coding and domain is None)
275
+ )
276
+
277
+ # 目前置信度仅基于规则分数,不是统计概率
278
+ return {
279
+ "skill": skill,
280
+ "knowledge": knowledge,
281
+ "domain": domain,
282
+ "confidence": confidence,
283
+ "reasons": reasons,
284
+ "needs_confirmation": needs_confirmation,
285
+ }
286
+
287
+
288
+ def show_result(result):
289
+ """展示推荐结果。"""
290
+
291
+ print("\n========== ZKCoding 推荐结果 ==========")
292
+ print("Skill:", result["skill"] or "暂无")
293
+ print(
294
+ "Knowledge:",
295
+ "、".join(result["knowledge"]) or "暂无推荐",
296
+ )
297
+ print("匹配领域:", result["domain"] or "未识别")
298
+ print("规则匹配强度:", result["confidence"])
299
+
300
+ print("\n推荐原因:")
301
+ for reason in result["reasons"]:
302
+ print("-", reason)
303
+
304
+ print("\n注意:建议用户确认后再执行。")
305
+ print("=======================================\n")
306
+
307
+
308
+ def run_tests():
309
+ """简单自测,不调用任何模型。"""
310
+
311
+ cases = [
312
+ (
313
+ "帮我用 Java 开发 ERP 采购报价系统",
314
+ "java_erp",
315
+ "erp",
316
+ ),
317
+ (
318
+ "帮我写一个 WMS 库存流水接口",
319
+ "java_wms",
320
+ "wms",
321
+ ),
322
+ (
323
+ "帮我开发设备巡检工单系统",
324
+ "java_device",
325
+ "device",
326
+ ),
327
+ (
328
+ "帮我开发一个网页计算器",
329
+ "java_calculator",
330
+ "calculator",
331
+ ),
332
+ (
333
+ "帮我开发一个 PDF 解析程序",
334
+ "java_pdf",
335
+ "pdf",
336
+ ),
337
+ (
338
+ "查询采购订单的付款状态",
339
+ "erp",
340
+ "erp",
341
+ ),
342
+ (
343
+ "查询仓库库存余额",
344
+ "wms",
345
+ "wms",
346
+ ),
347
+ (
348
+ "查询设备维修工单",
349
+ "device",
350
+ "device",
351
+ ),
352
+ ]
353
+
354
+ success = 0
355
+
356
+ for question, expected_skill, expected_domain in cases:
357
+ result = recommend(question)
358
+ expected_name = SKILLS[expected_skill]
359
+
360
+ correct = (
361
+ result["skill"] == expected_name
362
+ and result["domain"] == expected_domain
363
+ )
364
+
365
+ print("问题:", question)
366
+ print("结果:", "PASS" if correct else "FAIL")
367
+ print("推荐:", result["skill"], "\n")
368
+
369
+ if correct:
370
+ success += 1
371
+
372
+ print(f"自测结果: {success}/{len(cases)}")
373
+ return success == len(cases)
374
+
375
+
376
+ if __name__ == "__main__":
377
+ if "--self-test" in sys.argv:
378
+ if not run_tests():
379
+ sys.exit(1)
380
+ else:
381
+ while True:
382
+ question = input(
383
+ "请输入需求(输入 exit 退出):"
384
+ ).strip()
385
+
386
+ if question.lower() == "exit":
387
+ break
388
+
389
+ result = recommend(question)
390
+ show_result(result)