searxng/searx/engines/startpage.py
2013-11-09 18:39:20 +01:00

32 lines
997 B
Python

from urllib import urlencode
from lxml import html
from urlparse import urlparse
from cgi import escape
base_url = 'https://startpage.com/'
search_url = base_url+'do/search'
def request(query, params):
global search_url
query = urlencode({'q': query})[2:]
params['url'] = search_url
params['method'] = 'POST'
params['data'] = {'query': query}
return params
def response(resp):
global base_url
results = []
dom = html.fromstring(resp.content)
for result in dom.xpath('//div[@class="result"]'):
link = result.xpath('.//h3/a')[0]
url = link.attrib.get('href')
parsed_url = urlparse(url)
# TODO better google link detection
if parsed_url.netloc.find('www.google.com') >= 0:
continue
title = ' '.join(link.xpath('.//text()'))
content = escape(' '.join(result.xpath('.//p[@class="desc"]//text()')))
results.append({'url': url, 'title': title, 'content': content})
return results