searxng/searx/engines/bing_news.py

103 lines
3.3 KiB
Python

"""
Bing (News)
@website https://www.bing.com/news
@provide-api yes (http://datamarket.azure.com/dataset/bing/search),
max. 5000 query/month
@using-api no (because of query limit)
@results HTML (using search portal)
@stable no (HTML can change)
@parse url, title, content, publishedDate
"""
from urllib import urlencode
from cgi import escape
from lxml import html
from datetime import datetime, timedelta
from dateutil import parser
import re
from searx.engines.xpath import extract_text
# engine dependent config
categories = ['news']
paging = True
language_support = True
# search-url
base_url = 'https://www.bing.com/'
search_string = 'news/search?{query}&first={offset}'
# do search-request
def request(query, params):
offset = (params['pageno'] - 1) * 10 + 1
if params['language'] == 'all':
language = 'en-US'
else:
language = params['language'].replace('_', '-')
search_path = search_string.format(
query=urlencode({'q': query, 'setmkt': language}),
offset=offset)
params['cookies']['_FP'] = "ui=en-US"
params['url'] = base_url + search_path
return params
# get response from search-request
def response(resp):
results = []
dom = html.fromstring(resp.content)
# parse results
for result in dom.xpath('//div[@class="sn_r"]'):
link = result.xpath('.//div[@class="newstitle"]/a')[0]
url = link.attrib.get('href')
title = extract_text(link)
contentXPath = result.xpath('.//div[@class="sn_txt"]/div//span[@class="sn_snip"]')
content = escape(extract_text(contentXPath))
# parse publishedDate
publishedDateXPath = result.xpath('.//div[@class="sn_txt"]/div'
'//span[contains(@class,"sn_ST")]'
'//span[contains(@class,"sn_tm")]')
publishedDate = escape(extract_text(publishedDateXPath))
if re.match("^[0-9]+ minute(s|) ago$", publishedDate):
timeNumbers = re.findall(r'\d+', publishedDate)
publishedDate = datetime.now() - timedelta(minutes=int(timeNumbers[0]))
elif re.match("^[0-9]+ hour(s|) ago$", publishedDate):
timeNumbers = re.findall(r'\d+', publishedDate)
publishedDate = datetime.now() - timedelta(hours=int(timeNumbers[0]))
elif re.match("^[0-9]+ hour(s|), [0-9]+ minute(s|) ago$", publishedDate):
timeNumbers = re.findall(r'\d+', publishedDate)
publishedDate = datetime.now()\
- timedelta(hours=int(timeNumbers[0]))\
- timedelta(minutes=int(timeNumbers[1]))
elif re.match("^[0-9]+ day(s|) ago$", publishedDate):
timeNumbers = re.findall(r'\d+', publishedDate)
publishedDate = datetime.now() - timedelta(days=int(timeNumbers[0]))
else:
try:
publishedDate = parser.parse(publishedDate, dayfirst=False)
except TypeError:
publishedDate = datetime.now()
except ValueError:
publishedDate = datetime.now()
# append result
results.append({'url': url,
'title': title,
'publishedDate': publishedDate,
'content': content})
# return results
return results