| 123456789101112131415161718192021222324252627282930313233343536373839404142434445464748495051525354555657585960616263646566676869707172737475767778798081828384858687888990919293949596979899100101102103104105106107108109110111112113114115116117118119120121122123124125126127128129130131132133134135136137138139140141142143144145146147148149150151152153154155156157158159160161162163164165166167168169170171172173174175176177178179180181182183184185186187188189190191192193194195196197198199200201202203204205206207208209210211212213214215216217218219220221 |
- import streamlit as st
- import nltk
- import numpy as np
- import pandas as pd
- from sentence_transformers import SentenceTransformer, CrossEncoder
- from auth import login_to_wiki
- from scraping import fetch_all_pages_content
- from chunking import semantic_chunking, model
- from embedding_index import build_faiss_index, search_faiss
- from qa_chain import get_qa_chain
- import warnings
- import os
- import json
- warnings.filterwarnings("ignore")
- os.environ["STREAMLIT_SERVER_RUN_ON_SAVE"] = "false"
- # Pour l'affichage du niveau de pertinence
- def get_relevance_label(score):
- if score >= 2.5:
- return "✅ bon"
- elif score >= -2:
- return "⚠️ moyen"
- else:
- return "❌ faible"
- # Fonction principale de QA avec reranking et affichage complet
- 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]
- # Console output
- 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)
- print("\n🎯 Résultats re-rankés (top 10) :")
- print(df.head(10).to_string(index=True))
- # Sélection du contexte
- top_urls = [r["url"] for r in reranked]
- unique_top_urls = list(set(top_urls))
- if len(unique_top_urls) == 1:
- st.info("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])
- # Template
- template = """
- Tu es un assistant intelligent qui répond uniquement à partir du texte ci-dessous.
- Ne réponds que si la réponse est clairement présente. Si tu ne sais pas, dis "Je ne sais pas".
- CONTEXTE :
- {context}
- QUESTION :
- {question}
- RÉPONSE :
- """
- # Génération de la réponse
- result = qa_chain.run(context=full_context, question=question)
- # Affichage Streamlit
- st.write("### 🤖 Réponse")
- st.write(result)
- # Affichage terminal : prompt complet
- print("\n====================== PROMPT UTILISÉ ======================")
- print(template.format(context=full_context, question=question))
- print(result)
- print("============================================================\n")
- def main():
- st.title("🔎 Recherche Wiki")
- nltk.download('punkt')
- user = st.text_input("Entrez votre nom d'utilisateur")
- pwd = st.text_input("Entrez votre mot de passe", type="password")
- if st.button("Se connecter et charger données"):
- if login_to_wiki(user, pwd):
- st.success("Connexion réussie.")
- 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",
- "https://wiki.intersel.fr/doku.php?id=intersel:php",
- "https://wiki.intersel.fr/doku.php?id=produit:developpement:livinweb:documentations:admin_manual:start",
- "https://wiki.intersel.fr/doku.php?id=intersel:resume_annuaire",
- "https://wiki.intersel.fr/doku.php?id=technique:developpement:developpement_phplist:start",
- "https://wiki.intersel.fr/doku.php?id=produit:developpement:livinweb:projets:integration_metronic:start",
- "https://wiki.intersel.fr/doku.php?id=produit:developpement:livinweb:documentations:dev_manual:start"
- ]
- with st.spinner("Récupération et traitement des pages Wiki..."):
- url_texts = fetch_all_pages_content(wiki_urls)
- # Sauvegarde des contenus bruts de chaque page dans un fichier JSON lisible
- raw_pages = [{"url": url, "text": text.strip()} for url, text in url_texts.items()]
- with open("raw_pages_pretty.json", "w", encoding="utf-8") as f:
- json.dump(raw_pages, f, ensure_ascii=False, indent=2)
- print("📄 Contenu brut des pages enregistré dans 'raw_pages_pretty.json'")
- all_chunks = []
- chunk_urls = []
- for url, text in url_texts.items():
- sentences = nltk.sent_tokenize(text, language='french')
-
- print("\n" + "="*80)
- print(f"📄 Traitement de la page : {url}")
- print("="*80)
- print(f"🔢 Nombre de phrases détectées : {len(sentences)}")
- print(sentences)
- chunks = semantic_chunking(sentences)
- for chunk in chunks:
- all_chunks.append(chunk)
- chunk_urls.append(url)
- print("Nombre de mots total :", sum(len(s.split()) for s in sentences))
- print("Chunks générés :", len(chunks))
- print("\n🧩 Détail des chunks générés :")
- for i, chunk in enumerate(chunks):
- print("=" * 60)
- print(f"📌 Chunk {i+1}/{len(chunks)} | 📝 {len(chunk.split())} mots")
- print("-" * 60)
- print(chunk.strip())
- print("=" * 60 + "\n")
- 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')
- print("\n✅ Construction de l'index FAISS...")
- index = build_faiss_index(chunk_embeddings)
- print("✅ Index FAISS construit avec succès")
- chunk_data = [
- {
- "url": chunk_urls[i],
- "chunk_index": i + 1,
- "nb_mots": len(all_chunks[i].split()),
- "chunk": all_chunks[i]
- }
- for i in range(len(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'")
- cross_encoder = CrossEncoder('cross-encoder/ms-marco-MiniLM-L-6-v2')
- qa_chain = get_qa_chain()
- st.session_state['model'] = model
- st.session_state['cross_encoder'] = cross_encoder
- st.session_state['index'] = index
- st.session_state['all_chunks'] = all_chunks
- st.session_state['chunk_urls'] = chunk_urls
- st.session_state['qa_chain'] = qa_chain
- else:
- st.error("Échec de la connexion, veuillez vérifier vos identifiants.")
- if 'model' in st.session_state:
- question = st.text_input("❓ Posez votre question sur le Wiki")
- if st.button("Chercher"):
- ask_question_with_reranking(
- question,
- st.session_state['model'],
- st.session_state['cross_encoder'],
- st.session_state['index'],
- st.session_state['all_chunks'],
- st.session_state['chunk_urls'],
- st.session_state['qa_chain']
- )
- if __name__ == "__main__":
- main()
|