維基百科API知識圖譜RAG技術實作與評估

import textwrap
import wiki

# 定義主題
topic = "Marketing"

# 檢索維基百科頁面
page = wiki.page(topic)

if page.exists() == True:
    print("頁面存在：", page.exists())
    summary = page.summary
    nbt = nb_tokens(summary)
    print("Token數量：", nbt)
    
    # 使用textwrap包裝摘要文字
    wrapped_text = textwrap.fill(summary, width=60)
    print(wrapped_text)
else:
    print("頁面不存在")

# 取得頁面的完整URL
print(page.fullurl)

# 取得頁面中的所有連結
links = page.links
urls = []
counter = 0

for link in links:
    try:
        counter += 1
        print(f"連結 {counter}：{link}")
        summary = wiki.page(link).summary
        print(f"連結：{link}")
        print(wiki.page(link).fullurl)
        urls.append(wiki.page(link).fullurl)
        print(f"摘要：{summary}")
        
        if counter >= maxl:
            break
    except:
        # 忽略不存在的頁面
        pass

print(counter)
print(urls)

from datetime import datetime

# 準備寫入參照檔案
fname = filename + "_citations.txt"
with open(fname, "w") as file:
    file.write(f"參照。維基百科，自由的百科全書。\n")
    file.write("根頁面：" + page.fullurl + "\n")
    counter = 0
    
    for link in links:
        try:
            counter += 1
            summary = wiki.page(link).summary
            file.write(f"連結 {counter}：{link}\n")
            file.write(wiki.page(link).fullurl + "\n")
            file.write(f"摘要：{summary}\n")
            
            if counter >= maxl:
                break
        except:
            pass

# 寫入URL檔案
ufname = filename + "_urls.txt"
with open(ufname, 'w') as file:
    for url in urls:
        file.write(url + '\n')
    print("URL已寫入urls.txt")

from llama_index.core import KnowledgeGraphIndex
import time

# 開始計時
start_time = time.time()

# 建立知識圖譜索引
graph_index = KnowledgeGraphIndex.from_documents(
    documents,
    max_triplets_per_chunk=2,
    include_embeddings=True,
)

# 停止計時
end_time = time.time()
# 計算並列印執行時間
elapsed_time = end_time - start_time
print(f"索引建立時間：{elapsed_time:.4f} 秒")
print(type(graph_index))

索引建立時間：371.9844 秒
<class 'llama_index.core.indices.knowledge_graph.base.KnowledgeGraphIndex'>

# 相似度前k個
k = 3
# 溫度
temp = 0.1
# 輸出長度
mt = 1024

graph_query_engine = graph_index.as_query_engine(similarity_top_k=k, temperature=temp, num_output=mt)

from pyvis.network import Network

g = graph_index.get_networkx_graph()
net = Network(notebook=True, cdn_resources="in_line", directed=True)
net.from_nx(g)

# 設定節點和邊的屬性
for node in net.nodes:
    node['color'] = 'lightgray'
    node['size'] = 10
for edge in net.edges:
    edge['color'] = 'black'
    edge['width'] = 1

fgraph = "Knowledge_graph_" + graph_name + ".html"
net.write_html(fgraph)
print(fgraph)

response = graph_query_engine.query(user_input)

from sentence_transformers import SentenceTransformer
from sklearn.metrics.pairwise import cosine_similarity

model = SentenceTransformer('all-MiniLM-L6-v2')

def calculate_cosine_similarity_with_embeddings(text1, text2):
    embeddings1 = model.encode(text1)
    embeddings2 = model.encode(text2)
    similarity = cosine_similarity([embeddings1], [embeddings2])
    return similarity[0][0]

user_query = "Which experts are often associated with marketing theory?"
response = execute_query(user_query)

for idx, node_with_score in enumerate(response.source_nodes):
    similarity_score3 = calculate_cosine_similarity_with_embeddings(text1, text2)
    # ...

rscores = []
scores = []

for _ in range(10):
    user_query = # ...
    response = execute_query(user_query)
    similarity_score3 = calculate_cosine_similarity_with_embeddings(text1, text2)
    scores.append(similarity_score3)
    human_feedback = # ...
    rscores.append(human_feedback)

import numpy as np

# 餘弦相似度分數
scores = [0.808918, 0.720165, 0.7599532, 0.8513956, 0.5457667, 0.696391, 0.904, 0.2, 0.2, 0.9]

# 計算評估指標
mean_score = np.mean(scores)
median_score = np.median(scores)
std_deviation = np.std(scores)
variance = np.var(scores)
min_score = np.min(scores)
max_score = np.max(scores)
range_score = max_score - min_score
percentile_25 = np.percentile(scores, 25)
percentile_75 = np.percentile(scores, 75)
iqr = percentile_75 - percentile_25

print("平均值：", mean_score)
print("中位數：", median_score)
print("標準差：", std_deviation)
print("變異數：", variance)
print("最小值：", min_score)
print("最大值：", max_score)
print("範圍：", range_score)
print("25百分位數：", percentile_25)
print("75百分位數：", percentile_75)
print("四分位距（IQR）：", iqr)