Spaces:
Paused
Paused
File size: 3,499 Bytes
9acd20c | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 | import time
from sentence_transformers import SentenceTransformer
# 实验参数
BUDGET_BYTES = 8000
MODEL_NAME = "intfloat/multilingual-e5-small"
query = "如何实现 Java ERP 采购订单审批接口?"
# 模拟知识库返回顺序
documents = [
{
"id": "A",
"title": "Java 开发规范",
"text": "Java 项目应使用清晰的类名和方法名。" * 65,
"required": False,
},
{
"id": "B",
"title": "ERP 采购基础介绍",
"text": "ERP 系统用于管理采购订单及供应商信息。" * 55,
"required": False,
},
{
"id": "C",
"title": "较长的采购背景资料",
"text": "企业采购业务包含多个管理流程和操作步骤。" * 85,
"required": False,
},
{
"id": "D",
"title": "采购审批核心规则",
"text": (
"采购订单提交后必须经过审批,"
"审批通过后才能执行后续采购操作。"
),
"required": True,
},
]
def size_bytes(doc):
return len(str(doc).encode("utf-8"))
# A:原版遇到超限直接结束
def select_break(docs):
selected = []
used = 0
for doc in docs:
size = size_bytes(doc)
if used + size > BUDGET_BYTES:
break
selected.append(doc)
used += size
return selected
# B:跳过过长文档,继续检查
def select_continue(docs):
selected = []
used = 0
for doc in docs:
size = size_bytes(doc)
if used + size > BUDGET_BYTES:
continue
selected.append(doc)
used += size
return selected
# 加载你已经下载的 Embedding 模型
print("加载语义模型...")
model = SentenceTransformer(MODEL_NAME)
# 预先计算文档向量
passages = [
"passage: " + doc["title"] + " " + doc["text"]
for doc in documents
]
doc_vectors = model.encode(
passages,
normalize_embeddings=True,
)
# C:语义重排序 + 预算
def select_semantic(docs):
query_vector = model.encode(
"query: " + query,
normalize_embeddings=True,
)
scores = doc_vectors @ query_vector
ranked = sorted(
zip(docs, scores),
key=lambda item: float(item[1]),
reverse=True,
)
selected = []
used = 0
for doc, score in ranked:
size = size_bytes(doc)
if used + size > BUDGET_BYTES:
continue
selected.append(doc)
used += size
return selected
def evaluate(name, func):
start = time.perf_counter()
selected = func(documents)
elapsed = (time.perf_counter() - start) * 1000
used = sum(size_bytes(doc) for doc in selected)
has_required = any(
doc["required"] for doc in selected
)
print(f"\n【{name}】")
print("保留文档:", [d["id"] for d in selected])
print("使用字节:", used)
print("核心审批规则是否保留:", has_required)
print("处理耗时(ms):", round(elapsed, 3))
evaluate("A 原版 break", select_break)
evaluate("B 改进 continue", select_continue)
evaluate("C 语义重排序", select_semantic)
# D:直接保留前两条检索结果
def select_top2(docs):
selected = []
used = 0
for doc in docs[:2]:
size = size_bytes(doc)
if used + size > BUDGET_BYTES:
continue
selected.append(doc)
used += size
return selected
evaluate("D Top-2 精简", select_top2) |