pierre Brault
imit
3ff674d
raw
history blame contribute delete
5.08 kB
from datetime import datetime
import warnings
from typing import List
from pickle import Unpickler
import re
from bs4 import BeautifulSoup
from groq import Groq
from cohere import Client
from numpy.typing import NDArray
from numpy import array
from gossip_semantic_search.models import Article, Answer
from gossip_semantic_search.constant import (AUTHOR_KEY, TITLE_KEY, LINK_KEY, DESCRIPTION_KEY,
PUBLICATION_DATE_KEY, CONTENT_KEY, LLAMA_70B_MODEL,
DATE_FORMAT, EMBEDING_MODEL)
from gossip_semantic_search.prompts import (generate_question_prompt,
generate_context_retriver_prompt)
def xml_to_dict(element):
result = {}
for child in element:
child_dict = xml_to_dict(child)
if child.tag in result:
if isinstance(result[child.tag], list):
result[child.tag].append(child_dict)
else:
result[child.tag] = [result[child.tag], child_dict]
else:
result[child.tag] = child_dict
if element.text and element.text.strip():
result = element.text.strip()
return result
def sanitize_html_content(html_content):
soup = BeautifulSoup(html_content, 'html.parser')
for a in soup.find_all('a'):
a.unwrap()
for tag in soup.find_all(['em', 'strong']):
tag.unwrap()
for blockquote in soup.find_all('blockquote'):
blockquote.extract()
cleaned_text = re.sub(r'\s+', ' ', soup.get_text()).strip()
return cleaned_text
def article_raw_to_article(raw_article) -> Article:
return Article(
author = raw_article[AUTHOR_KEY],
title = raw_article[TITLE_KEY],
link = raw_article[LINK_KEY],
description = raw_article[DESCRIPTION_KEY],
published_date = datetime.strptime(
raw_article[PUBLICATION_DATE_KEY],
DATE_FORMAT
),
content = sanitize_html_content(raw_article[CONTENT_KEY])
)
def generates_questions(context: str,
nb_questions: int,
client: Groq) -> List[str]:
completion = client.chat.completions.create(
model=LLAMA_70B_MODEL,
messages=[
{
"role": "user",
"content": generate_question_prompt(context, nb_questions)
},
],
temperature=1,
max_tokens=1024,
top_p=1,
stream=True,
stop=None,
)
questions_str = "".join(chunk.choices[0].delta.content or "" for chunk in completion)
try:
questions = re.findall(r'([^?]*\?)', questions_str)
questions = [question.strip()[3:] for question in questions]
except IndexError:
warnings.warn(f"no question found. \n"
f"string return: {questions_str}")
return []
if len(questions) != nb_questions:
warnings.warn(f"Expected {nb_questions} questions, but found "
f"{len(questions)}. {', '.join(questions)}", UserWarning)
return questions
def choose_context_and_answer_questions(articles: List[Article],
query:str,
generative_client) -> Answer:
for article in articles:
completion = generative_client.chat.completions.create(
model=LLAMA_70B_MODEL,
messages=[
{
"role": "user",
"content": generate_context_retriver_prompt(query, article.content)
},
],
temperature=1,
max_tokens=1024,
top_p=1,
stream=True,
stop=None,
)
answer = "".join(chunk.choices[0].delta.content or "" for chunk in completion)
pattern = r"answer_in_text\s*=\s*(.*?),"
# Appliquer la regex
match = re.search(pattern, answer)
if match:
if match.group(1) == "True":
pattern = r"answer\s*=\s*(.*)"
match = re.search(pattern, answer)
if match:
answer_value = match.group(1)[1:-2]
return Answer(
answer = answer_value,
link = f"{article.link}",
content = f"{article.content}"
)
return Answer(
answer = "incapable de générer une reponse",
link = f"{articles[0].link}",
content = f"{articles[0].content}"
)
def embed_content(contexts:List[str],
client: Client) -> NDArray:
return array(client.embed(
model=EMBEDING_MODEL,
texts=contexts,
input_type='classification',
truncate='NONE'
).embeddings)
class CustomUnpickler(Unpickler):
def find_class(self, module, name):
if module == 'models':
return Article # Renvoie une classe de remplacement
return super().find_class(module, name)