| 123456789101112131415161718192021222324252627282930313233343536373839404142434445464748495051525354555657585960616263646566676869707172737475767778798081828384858687888990919293949596979899100101102103104105106107108109110111112113114115116117118119120121122123124125126127128129130131132133134135136137138139140141142143144145146147148149150151152153154155156157158159160161162163164165166167168169170171172173174175176177178179180181182183184185186187188189190191192193194195196197198199200201202203204205206207208209210211212213214215216217218219220221222223224225226227228229230231232233234235236237238239240241242243244245246247248249250251252253254255256257258259260261262263264265266267268269270271272273274275276277278279280281282283284285286287288289290291292293294295296297298299300301302303304305306307308309310311312313314315316317318319320321322323324325326 |
- import requests
- from bs4 import BeautifulSoup
- from langchain_community.llms import Ollama
- from langchain.prompts import PromptTemplate
- from langchain.chains import LLMChain
- import warnings
- import nltk
- import spacy
- from nltk.corpus import stopwords
- from sentence_transformers import SentenceTransformer, CrossEncoder
- import torch
- import pandas as pd
- import faiss
- import numpy as np
- from langchain.text_splitter import RecursiveCharacterTextSplitter
- import json
- from urllib.parse import urljoin, urlparse
- from bs4 import SoupStrainer
- from sklearn.metrics.pairwise import cosine_similarity
- import getpass
- #username = input("Nom d'utilisateur : ")
- #password = getpass.getpass("Mot de passe : ")
- #login_to_wiki(username, password)
- warnings.filterwarnings("ignore")
- nltk.download('punkt')
- nltk.download('stopwords')
- stop_words = set(stopwords.words('french'))
- nlp = spacy.load('fr_core_news_sm')
- session = requests.Session()
- def login_to_wiki(username, password):
- login_url = "https://wiki.intersel.fr/doku.php?id=start&do=login"
- resp = session.get(login_url)
- soup = BeautifulSoup(resp.text, "html.parser")
- sectok_input = soup.find("input", {"name": "sectok"})
- if not sectok_input:
- print("❌ Impossible de trouver le token CSRF.")
- return False
- login_data = {
- "u": username,
- "p": password,
- "do": "login",
- "sectok": sectok_input.get("value")
- }
- headers = {
- "User-Agent": "Mozilla/5.0",
- "Referer": login_url
- }
- login_resp = session.post(login_url, data=login_data, headers=headers)
- if "logout" in login_resp.text.lower():
- print("✅ Connexion réussie.")
- return True
- else:
- print("❌ Échec de la connexion.")
- return False
- def extract_visible_text(html):
- soup = BeautifulSoup(html, "html.parser")
- for tag in soup(["script", "style", "header", "footer", "nav"]):
- tag.decompose()
- text = soup.get_text(separator="\n")
- lines = [line.strip() for line in text.splitlines() if line.strip()]
- return "\n".join(lines)
- #def clean_text(text):
- # doc = nlp(text)
- # tokens = [token.text for token in doc if token.text.lower() not in stop_words and token.text.strip() != ""]
- # return ' '.join(tokens)
- def chunk_by_titles(text):
- """
- Découpe le texte en chunks basés sur les titres DokuWiki (=, ==, === etc).
- Chaque chunk correspond à une section commençant par un titre.
- """
- chunks = []
- current_chunk = []
- lines = text.split('\n')
- for line in lines:
- # Si la ligne commence par un signe '=' => début d'un nouveau chunk
- if line.strip().startswith('='):
- # Si on a déjà un chunk en cours, on l'ajoute à la liste
- if current_chunk:
- chunks.append('\n'.join(current_chunk).strip())
- current_chunk = []
- current_chunk.append(line)
- else:
- current_chunk.append(line)
- # Ajouter le dernier chunk
- if current_chunk:
- chunks.append('\n'.join(current_chunk).strip())
- return chunks
- def fetch_all_pages_content(urls):
- url_texts = {}
- for url in urls:
- try:
- edit_url = url + "&do=edit" if "?" in url else url + "?do=edit"
- response = session.get(edit_url)
- if response.status_code == 200:
- soup = BeautifulSoup(response.text, "html.parser")
- textarea = soup.find("textarea", {"name": "wikitext"})
- if textarea:
- raw_text = textarea.text
- url_texts[url] = raw_text
- else:
- print(f"⚠️ Pas de balise <textarea> trouvée pour {edit_url}.")
- else:
- print(f"⚠️ Erreur d'accès à {edit_url} : code {response.status_code}")
- except Exception as e:
- print(f"❌ Exception pour {url}: {e}")
- return url_texts
- def build_faiss_index(embeddings):
- dimension = embeddings.shape[1]
- index = faiss.IndexFlatIP(dimension)
- faiss.normalize_L2(embeddings)
- index.add(embeddings)
- return index
- def search_faiss(index, query_embedding, top_k=10):
- faiss.normalize_L2(query_embedding)
- distances, indices = index.search(query_embedding, top_k)
- return distances[0], indices[0]
- def get_relevance_label(score):
- if score >= 2.5:
- return "✅ bon"
- elif score >= 0:
- return "⚠️ moyen"
- else:
- return "❌ faible"
- def ask_question_with_reranking(question, model, cross_encoder, index, chunks, chunk_urls, qa_chain, faiss_top_k=30, rerank_top_k=10):
- question_embedding = model.encode([question], convert_to_tensor=False)
- question_embedding = np.array(question_embedding).astype('float32')
- if question_embedding.ndim == 1:
- question_embedding = np.expand_dims(question_embedding, axis=0)
- distances, indices = search_faiss(index, question_embedding, top_k=faiss_top_k)
- candidates = []
- for idx, dist in zip(indices, distances):
- candidates.append({
- "chunk": chunks[idx],
- "url": chunk_urls[idx],
- "index": idx,
- "faiss_score": float(dist)
- })
- # Reranking
- pairs = [[question, c["chunk"]] for c in candidates]
- rerank_scores = cross_encoder.predict(pairs)
- for i in range(len(candidates)):
- candidates[i]["rerank_score"] = float(rerank_scores[i])
- reranked = sorted(candidates, key=lambda x: x["rerank_score"], reverse=True)[:rerank_top_k]
- print("\n🔍 Résultats re-rankés :")
- df = pd.DataFrame([{
- "Index": r["index"],
- "URL": r["url"],
- "Score CrossEncoder": round(r["rerank_score"], 4),
- "Chunk": r["chunk"][:100] + "..." if len(r["chunk"]) > 100 else r["chunk"]
- } for r in reranked])
- df["Pertinence"] = df["Score CrossEncoder"].apply(get_relevance_label)
- pd.set_option('display.max_colwidth', None)
- df_reranked = df.copy()
- # Affichage formaté ligne par ligne
- for i, row in df_reranked.iterrows():
- print(f"{i:>3} {row['URL']:<70} {row['Score CrossEncoder']:.4f} {row['Pertinence']}")
- print(f"{row['Chunk'][:150]}...\n")
- # Tableau propre
- colonnes_a_afficher = ['Index', 'URL', 'Score CrossEncoder', 'Pertinence', 'Chunk']
- df_affichage = df_reranked[colonnes_a_afficher].copy()
- df_affichage.reset_index(drop=True, inplace=True)
- print("\n🎯 Résultats re-rankés (top 10) :")
- print(df_affichage.head(10).to_string(index=True))
- # Si tous les top chunks viennent d'une seule page
- top_urls = [r["url"] for r in reranked]
- unique_top_urls = list(set(top_urls))
- if len(unique_top_urls) == 1:
- print("ℹ️ Tous les meilleurs chunks viennent de la même page. Utilisation de l'ensemble de la page.")
- full_context = "\n\n".join([
- chunks[i] for i in range(len(chunks)) if chunk_urls[i] == unique_top_urls[0]
- ])
- else:
- full_context = "\n\n".join([r["chunk"] for r in reranked])
- result = qa_chain.run(context=full_context, question=question)
- print(f"\n🤖 Réponse :\n{result}\n")
- def semantic_chunking(sentences, similarity_threshold=0.5):
- chunks = []
- current_chunk = []
- current_embeddings = []
-
- for sentence in sentences:
- embedding = model.encode(sentence)
- if not current_embeddings:
- current_chunk.append(sentence)
- current_embeddings.append(embedding)
- else:
- similarities = cosine_similarity([embedding], current_embeddings)[0]
- if max(similarities) > similarity_threshold:
- current_chunk.append(sentence)
- current_embeddings.append(embedding)
- else:
- chunks.append(" ".join(current_chunk))
- current_chunk = [sentence]
- current_embeddings = [embedding]
- if current_chunk:
- chunks.append(" ".join(current_chunk))
- return chunks
- if __name__ == "__main__":
- model = SentenceTransformer('sentence-transformers/paraphrase-multilingual-MiniLM-L12-v2')
- if login_to_wiki("nadine.elleuch", "witpowgeu!75"):
- wiki_urls = [
- "https://wiki.intersel.fr/doku.php?id=intersel:presentation_equipe",
- "https://wiki.intersel.fr/doku.php?id=intersel:presentation",
- "https://wiki.intersel.fr/doku.php?id=intersel:guides:entretiens:stagiaires",
- "https://wiki.intersel.fr/doku.php?id=projets:samsic:anticipia:start"
- ]
- url_texts = fetch_all_pages_content(wiki_urls)
- all_chunks = []
- chunk_urls = []
- for url, text in url_texts.items():
- print(f"\n--- Traitement de la page {url} ---")
- # Split the text into sentences
- sentences = nltk.sent_tokenize(text, language='french')
- # Semantic chunking using similarity
- chunks = semantic_chunking(sentences)
- print(f"Nombre de chunks : {len(chunks)}")
- for i, chunk in enumerate(chunks):
- print(f"Chunk {i+1} ({len(chunk.split())} mots): {chunk[:100]}{'...' if len(chunk) > 100 else ''}")
- all_chunks.append(chunk)
- chunk_urls.append(url)
- # Prepare chunk metadata
- chunk_data = [
- {"url": url_, "chunk_index": i + 1, "chunk": chunk}
- for i, (url_, chunk) in enumerate(zip(chunk_urls, all_chunks))
- ]
- with open("chunks_pretty.json", "w", encoding="utf-8") as f:
- json.dump(chunk_data, f, ensure_ascii=False, indent=2)
- print("📁 Chunks enregistrés dans 'chunks_pretty.json'")
- # Load reranker
- cross_encoder = CrossEncoder('cross-encoder/ms-marco-MiniLM-L-6-v2')
- # Encode the chunks
- print("\n✅ Génération des embeddings des chunks...")
- chunk_embeddings = model.encode(all_chunks, convert_to_tensor=False, show_progress_bar=True)
- chunk_embeddings = np.array(chunk_embeddings).astype('float32')
- # Build FAISS index
- print("\n✅ Construction de l'index FAISS...")
- index = build_faiss_index(chunk_embeddings)
- print("✅ Index FAISS construit avec succès.")
- # Define prompt
- prompt_template = PromptTemplate(
- input_variables=["context", "question"],
- template="""
- Tu es un assistant expert du wiki interne d'Intersel. Réponds uniquement à partir du CONTEXTE donné.
- Si aucune information pertinente n'est disponible, dis "Je ne sais pas." sans inventer.
- CONTEXTE :
- {context}
- QUESTION :
- {question}
- RÉPONSE :
- """
- )
- llm = Ollama(model="llama3")
- qa_chain = LLMChain(llm=llm, prompt=prompt_template)
- while True:
- question = input("❓ Posez votre question sur le Wiki ('exit' pour quitter) :\n> ")
- if question.lower() in ["exit", "quit"]:
- break
- ask_question_with_reranking(
- question, model, cross_encoder, index,
- all_chunks, chunk_urls, qa_chain
- )
|