Spaces:
Running
Running
Update src/Algorithms/vector_search.py
Browse files- src/Algorithms/vector_search.py +20 -76
src/Algorithms/vector_search.py
CHANGED
|
@@ -1,101 +1,45 @@
|
|
| 1 |
"""
|
| 2 |
-
MODULE: VECTOR SEARCH ENGINE (
|
| 3 |
-
============================================
|
| 4 |
-
Amélioration : Le vecteur d'un nœud inclut désormais ses voisins et ses relations.
|
| 5 |
"""
|
| 6 |
import faiss
|
| 7 |
import numpy as np
|
| 8 |
from sentence_transformers import SentenceTransformer
|
| 9 |
-
import networkx as nx
|
| 10 |
|
| 11 |
class SemanticIndex:
|
| 12 |
def __init__(self, model_name='all-MiniLM-L6-v2'):
|
| 13 |
-
# Chargement du modèle (rapide sur CPU)
|
| 14 |
self.model = SentenceTransformer(model_name)
|
| 15 |
self.index = None
|
| 16 |
self.uris = []
|
| 17 |
-
self.
|
| 18 |
-
self.is_ready = False
|
| 19 |
|
| 20 |
-
def
|
| 21 |
-
"""
|
| 22 |
-
|
| 23 |
-
Texte = Attributs du Nœud + Relations Sortantes (Voisins)
|
| 24 |
-
"""
|
| 25 |
-
print("⏳ [VECTOR] Génération des embeddings contextuels...")
|
| 26 |
-
corpus = []
|
| 27 |
-
self.uris = []
|
| 28 |
-
self.metadatas = []
|
| 29 |
|
| 30 |
-
|
| 31 |
-
|
| 32 |
-
|
| 33 |
-
# On prend toutes les valeurs textuelles du nœud
|
| 34 |
-
internal_attrs = [
|
| 35 |
-
str(v) for k, v in data.items()
|
| 36 |
-
if k not in ['group', 'label', 'color', 'shape', 'x', 'y', 'size', 'title']
|
| 37 |
-
and v and str(v).lower() != 'nan'
|
| 38 |
-
]
|
| 39 |
-
base_text = f"{data.get('group', '')} {data.get('label', '')} {' '.join(internal_attrs)}"
|
| 40 |
|
| 41 |
-
# 2. Texte du Voisinage (Relations Sortantes)
|
| 42 |
-
# "travaille_chez X", "habite_ville Y"
|
| 43 |
-
neighbors_text = []
|
| 44 |
-
for neighbor in G.successors(node):
|
| 45 |
-
# Récupération du type de lien (ex: 'habite_ville')
|
| 46 |
-
edge_data = G.get_edge_data(node, neighbor)
|
| 47 |
-
rel_label = edge_data.get('label', 'lié_à')
|
| 48 |
-
|
| 49 |
-
# Récupération du nom du voisin (ex: 'Dakar')
|
| 50 |
-
neighbor_label = G.nodes[neighbor].get('label', str(neighbor))
|
| 51 |
-
|
| 52 |
-
# On ajoute la phrase de contexte
|
| 53 |
-
neighbors_text.append(f"{rel_label} {neighbor_label}")
|
| 54 |
-
|
| 55 |
-
# 3. Fusion (Soup)
|
| 56 |
-
# Ex: "Client Jean Dupont. habite_ville Dakar. secteur_act Commercial."
|
| 57 |
-
full_context = f"{base_text}. {' '.join(neighbors_text)}"
|
| 58 |
-
|
| 59 |
-
# Nettoyage
|
| 60 |
-
full_context = full_context.lower().replace('_', ' ')
|
| 61 |
-
|
| 62 |
-
corpus.append(full_context)
|
| 63 |
-
self.uris.append(node)
|
| 64 |
-
|
| 65 |
-
# Méta-données pour l'affichage dans le chat
|
| 66 |
-
meta_desc = f"{data.get('group', 'Entité')} - {data.get('label', str(node))}"
|
| 67 |
-
self.metadatas.append(meta_desc)
|
| 68 |
-
|
| 69 |
-
if not corpus:
|
| 70 |
-
print("⚠️ Graphe vide, pas de vectorisation.")
|
| 71 |
-
return
|
| 72 |
-
|
| 73 |
-
# 4. Encodage Batch (Rapide)
|
| 74 |
embeddings = self.model.encode(corpus, show_progress_bar=True)
|
| 75 |
-
|
| 76 |
-
# 5. Création Index FAISS
|
| 77 |
dimension = embeddings.shape[1]
|
| 78 |
self.index = faiss.IndexFlatL2(dimension)
|
| 79 |
self.index.add(np.array(embeddings).astype('float32'))
|
| 80 |
|
| 81 |
-
|
| 82 |
-
print(f"✅ [VECTOR] Index FAISS Contextuel prêt : {len(self.uris)} entités.")
|
| 83 |
-
return len(self.uris)
|
| 84 |
-
|
| 85 |
-
def search(self, query, top_k=10):
|
| 86 |
-
"""Retourne les nœuds les plus proches sémantiquement"""
|
| 87 |
-
if not self.is_ready:
|
| 88 |
-
return []
|
| 89 |
|
| 90 |
-
|
| 91 |
-
|
|
|
|
|
|
|
| 92 |
|
| 93 |
results = []
|
| 94 |
-
for
|
| 95 |
-
if idx
|
|
|
|
| 96 |
results.append({
|
| 97 |
-
"uri":
|
| 98 |
-
"
|
| 99 |
-
"score": float(
|
| 100 |
})
|
| 101 |
return results
|
|
|
|
| 1 |
"""
|
| 2 |
+
MODULE: VECTOR SEARCH ENGINE (OG-RAG VERSION)
|
| 3 |
+
=============================================
|
|
|
|
| 4 |
"""
|
| 5 |
import faiss
|
| 6 |
import numpy as np
|
| 7 |
from sentence_transformers import SentenceTransformer
|
|
|
|
| 8 |
|
| 9 |
class SemanticIndex:
|
| 10 |
def __init__(self, model_name='all-MiniLM-L6-v2'):
|
|
|
|
| 11 |
self.model = SentenceTransformer(model_name)
|
| 12 |
self.index = None
|
| 13 |
self.uris = []
|
| 14 |
+
self.content_map = {} # Pour retrouver le texte du bloc
|
|
|
|
| 15 |
|
| 16 |
+
def build_from_ontology_blocks(self, blocks):
|
| 17 |
+
"""Vectorise les Hyper-blocs enrichis"""
|
| 18 |
+
print(f"⏳ [VECTOR] Indexation de {len(blocks)} Hyper-blocs...")
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 19 |
|
| 20 |
+
corpus = [b['text'] for b in blocks]
|
| 21 |
+
self.uris = [b['uri'] for b in blocks]
|
| 22 |
+
self.content_map = {b['uri']: b['text'] for b in blocks}
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 23 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 24 |
embeddings = self.model.encode(corpus, show_progress_bar=True)
|
|
|
|
|
|
|
| 25 |
dimension = embeddings.shape[1]
|
| 26 |
self.index = faiss.IndexFlatL2(dimension)
|
| 27 |
self.index.add(np.array(embeddings).astype('float32'))
|
| 28 |
|
| 29 |
+
print("✅ [VECTOR] Index OG-RAG prêt.")
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 30 |
|
| 31 |
+
def search(self, query, top_k=3):
|
| 32 |
+
if not self.index: return []
|
| 33 |
+
query_vec = self.model.encode([query]).astype('float32')
|
| 34 |
+
dist, indices = self.index.search(query_vec, top_k)
|
| 35 |
|
| 36 |
results = []
|
| 37 |
+
for idx in indices[0]:
|
| 38 |
+
if idx != -1:
|
| 39 |
+
uri = self.uris[idx]
|
| 40 |
results.append({
|
| 41 |
+
"uri": uri,
|
| 42 |
+
"text": self.content_map[uri],
|
| 43 |
+
"score": float(dist[0][0])
|
| 44 |
})
|
| 45 |
return results
|