Update app.py
Browse files
app.py
CHANGED
@@ -2,89 +2,284 @@ import time
|
|
2 |
import gradio as gr
|
3 |
from datasets import load_dataset
|
4 |
import pandas as pd
|
5 |
-
from sentence_transformers import SentenceTransformer
|
6 |
from sentence_transformers.quantization import quantize_embeddings
|
7 |
import faiss
|
8 |
from usearch.index import Index
|
|
|
|
|
|
|
|
|
9 |
|
10 |
-
# Load titles and texts
|
11 |
-
wikipedia_dataset = load_dataset("bourdoiscatie/wikipedia_fr_2022_250K", split="train", num_proc=4).select_columns(["title", "text", "wiki_id"])
|
12 |
|
|
|
|
|
|
|
|
|
|
|
|
|
13 |
def add_link(example):
|
14 |
example["title"] = '['+example["title"]+']('+'https://fr.wikipedia.org/wiki?curid='+str(example["wiki_id"])+')'
|
15 |
return example
|
16 |
wikipedia_dataset = wikipedia_dataset.map(add_link)
|
17 |
-
|
18 |
-
|
|
|
19 |
int8_view = Index.restore("wikipedia_fr_2022_250K_int8_usearch.index", view=True)
|
20 |
binary_index: faiss.IndexBinaryFlat = faiss.read_index_binary("wikipedia_fr_2022_250K_ubinary_faiss.index")
|
21 |
binary_ivf: faiss.IndexBinaryIVF = faiss.read_index_binary("wikipedia_fr_2022_250K_ubinary_ivf_faiss.index")
|
22 |
|
23 |
-
|
24 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
25 |
|
26 |
|
27 |
-
|
28 |
-
|
|
|
|
|
|
|
|
|
29 |
start_time = time.time()
|
30 |
-
query_embedding =
|
31 |
embed_time = time.time() - start_time
|
32 |
|
33 |
-
# 2. Quantize the query to ubinary
|
34 |
start_time = time.time()
|
35 |
query_embedding_ubinary = quantize_embeddings(query_embedding.reshape(1, -1), "ubinary")
|
36 |
quantize_time = time.time() - start_time
|
37 |
|
38 |
-
# 3. Search the binary index (either exact or approximate)
|
39 |
index = binary_ivf if use_approx else binary_index
|
40 |
start_time = time.time()
|
41 |
-
_scores, binary_ids = index.search(query_embedding_ubinary, top_k * rescore_multiplier)
|
42 |
binary_ids = binary_ids[0]
|
43 |
-
|
44 |
|
45 |
-
# 4. Load the corresponding int8 embeddings
|
46 |
start_time = time.time()
|
47 |
int8_embeddings = int8_view[binary_ids].astype(int)
|
48 |
load_time = time.time() - start_time
|
49 |
|
50 |
-
# 5. Rescore the top_k * rescore_multiplier using the float32 query embedding and the int8 document embeddings
|
51 |
start_time = time.time()
|
52 |
scores = query_embedding @ int8_embeddings.T
|
53 |
rescore_time = time.time() - start_time
|
54 |
|
55 |
-
#
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
56 |
start_time = time.time()
|
57 |
-
|
58 |
-
|
59 |
-
|
60 |
-
|
61 |
-
|
62 |
-
|
63 |
-
|
64 |
-
|
65 |
-
|
66 |
-
|
67 |
-
|
68 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
69 |
sort_time = time.time() - start_time
|
|
|
|
|
|
|
|
|
|
|
|
|
70 |
|
71 |
-
return df,
|
72 |
-
"Temps pour enchâsser la requête ": f"{embed_time:.4f} s",
|
73 |
-
"Temps pour la quantisation ": f"{quantize_time:.4f} s",
|
74 |
-
"Temps pour effectuer la recherche ": f"{search_time:.4f} s",
|
75 |
-
"Temps de chargement ": f"{load_time:.4f} s",
|
76 |
-
"Temps de rescorage ": f"{rescore_time:.4f} s",
|
77 |
-
"Temps pour trier les résustats ": f"{sort_time:.4f} s",
|
78 |
-
"Temps total pour la recherche ": f"{quantize_time + search_time + load_time + rescore_time + sort_time:.4f} s",
|
79 |
-
}
|
80 |
|
81 |
|
82 |
-
with gr.Blocks(title="Requêter Wikipedia
|
83 |
|
84 |
gr.Markdown(
|
85 |
"""
|
86 |
## Requêter Wikipedia en temps réel 🔍
|
87 |
-
|
88 |
Ce démonstrateur permet de requêter un corpus composé des 250K paragraphes les plus consultés du Wikipédia francophone.
|
89 |
Les résultats sont renvoyés en temps réel via un pipeline tournant sur un CPU 🚀
|
90 |
Nous nous sommes grandement inspirés du Space [quantized-retrieval](https://huggingface.co/spaces/sentence-transformers/quantized-retrieval) conçu par [Tom Aarsen](https://huggingface.co/tomaarsen) 🤗
|
@@ -102,23 +297,11 @@ Il n'est pas exclus d'ensuite utiliser une version plus récente de Wikipedia (o
|
|
102 |
</details>
|
103 |
|
104 |
<details><summary>2. Détails le pipeline</summary>
|
105 |
-
|
106 |
-
2. La requête est quantizée en binaire à l'aide de la fonction `quantize_embeddings` de la bibliothèque <a href="https://sbert.net/">SentenceTransformers</a>.
|
107 |
-
3. Un index binaire (250K <i>embeddings</i> binaires pesant 32MB de mémoire/espace disque) est requêté (en binaire si l'option approximative est sélectionnée, en int8 si l'option exacte est sélectionnée).
|
108 |
-
4. Les <i>n</i> textes demandés par l'utilisateur jugés les plus pertinents sont chargés à la volée à partir d'un index int8 sur disque (250K <i>embeddings</i> int8 ; 0 bytes de mémoire, 293MB d'espace disque).
|
109 |
-
5. Les <i>n</i> textes sont rescorés en utilisant la requête en float32 et les enchâssements en int8.
|
110 |
-
6. Les <i>n</i> premiers textes sont triés par score et affichés. Le "Score_paragraphe" correspond au score individuel de chaque paragraphe d'être pertinant vis-à-vis de la requête. Le "Score_article" correspond à la somme de tous les scores individuels des paragraphes issus d'un même article Wikipedia. L'objectif est alors de mettre en avant l'article source plutôt qu'un bout de texte le composant.
|
111 |
-
|
112 |
-
Ce processus est conçu pour être rapide et efficace en termes de mémoire : l'index binaire étant suffisamment petit pour tenir dans la mémoire et l'index int8 étant chargé en tant que vue pour économiser de la mémoire.
|
113 |
-
Au total, ce processus nécessite de conserver 1) le modèle en mémoire, 2) l'index binaire en mémoire et 3) l'index int8 sur le disque.
|
114 |
-
Avec une dimension de 1024, nous avons besoin de `1024 / 8 * num_docs` octets pour l'index binaire et de `1024 * num_docs` octets pour l'index int8.
|
115 |
-
|
116 |
-
C'est nettement moins cher que de faire le même processus avec des enchâssements en float32 qui nécessiterait `4 * 1024 * num_docs` octets de mémoire/espace disque pour l'index float32, soit 32x plus de mémoire et 4x plus d'espace disque.
|
117 |
-
De plus, l'index binaire est beaucoup plus rapide (jusqu'à 32x) à rechercher que l'index float32, tandis que le rescorage est également extrêmement efficace.
|
118 |
-
En conclusion, ce processus permet une recherche rapide, évolutive, peu coûteuse et efficace en termes de mémoire.
|
119 |
</details>
|
120 |
"""
|
121 |
)
|
|
|
122 |
with gr.Row():
|
123 |
with gr.Column(scale=75):
|
124 |
query = gr.Textbox(
|
@@ -129,7 +312,7 @@ En conclusion, ce processus permet une recherche rapide, évolutive, peu coûteu
|
|
129 |
use_approx = gr.Radio(
|
130 |
choices=[("Exacte", False), ("Approximative", True)],
|
131 |
value=True,
|
132 |
-
label="Type de recherche",
|
133 |
)
|
134 |
|
135 |
with gr.Row():
|
@@ -139,8 +322,7 @@ En conclusion, ce processus permet une recherche rapide, évolutive, peu coûteu
|
|
139 |
maximum=40,
|
140 |
step=1,
|
141 |
value=15,
|
142 |
-
label="Nombre de documents à
|
143 |
-
info="Recherche effectué via un bi-encodeur binaire",
|
144 |
)
|
145 |
with gr.Column(scale=2):
|
146 |
rescore_multiplier = gr.Slider(
|
@@ -149,17 +331,72 @@ En conclusion, ce processus permet une recherche rapide, évolutive, peu coûteu
|
|
149 |
step=1,
|
150 |
value=1,
|
151 |
label="Coefficient de rescorage",
|
152 |
-
info="
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
153 |
)
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
154 |
|
155 |
-
search_button = gr.Button(value="
|
156 |
|
157 |
output = gr.Dataframe(headers=["Score_article", "Score_paragraphe", "Titre", "Texte"], datatype="markdown")
|
158 |
-
json = gr.JSON()
|
159 |
-
|
160 |
-
query.submit(search, inputs=[query, top_k, rescore_multiplier, use_approx], outputs=[output, json])
|
161 |
-
search_button.click(search, inputs=[query, top_k, rescore_multiplier, use_approx], outputs=[output, json])
|
162 |
-
gr.Image("/file=catie(2).png", height=250,width=80, show_download_button=False)
|
163 |
|
|
|
|
|
|
|
164 |
demo.queue()
|
165 |
-
demo.launch(
|
|
|
2 |
import gradio as gr
|
3 |
from datasets import load_dataset
|
4 |
import pandas as pd
|
5 |
+
from sentence_transformers import SentenceTransformer, SparseEncoder
|
6 |
from sentence_transformers.quantization import quantize_embeddings
|
7 |
import faiss
|
8 |
from usearch.index import Index
|
9 |
+
import torch
|
10 |
+
import numpy as np
|
11 |
+
from collections import defaultdict
|
12 |
+
from scipy import stats
|
13 |
|
|
|
|
|
14 |
|
15 |
+
|
16 |
+
# Load titles, texts and pre-computed sparse embeddings
|
17 |
+
wikipedia_dataset = load_dataset("CATIE-AQ/wikipedia_fr_2022_250K", split="train", num_proc=4).select_columns(["title", "text", "wiki_id", "sparse_emb"])
|
18 |
+
|
19 |
+
|
20 |
+
# A function to make the titles of Wikipedia articles clickable in the final dataframe, so that you can consult the article on the website
|
21 |
def add_link(example):
|
22 |
example["title"] = '['+example["title"]+']('+'https://fr.wikipedia.org/wiki?curid='+str(example["wiki_id"])+')'
|
23 |
return example
|
24 |
wikipedia_dataset = wikipedia_dataset.map(add_link)
|
25 |
+
|
26 |
+
|
27 |
+
# Load the int8 and binary indices for dense retrieval
|
28 |
int8_view = Index.restore("wikipedia_fr_2022_250K_int8_usearch.index", view=True)
|
29 |
binary_index: faiss.IndexBinaryFlat = faiss.read_index_binary("wikipedia_fr_2022_250K_ubinary_faiss.index")
|
30 |
binary_ivf: faiss.IndexBinaryIVF = faiss.read_index_binary("wikipedia_fr_2022_250K_ubinary_ivf_faiss.index")
|
31 |
|
32 |
+
|
33 |
+
# Load models
|
34 |
+
dense_model = SentenceTransformer("OrdalieTech/Solon-embeddings-large-0.1")
|
35 |
+
sparse_model = SparseEncoder("CATIE-AQ/SPLADE_camembert-base_STS")
|
36 |
+
|
37 |
+
|
38 |
+
# Fusion methods
|
39 |
+
def reciprocal_rank_fusion(dense_results, sparse_results, k=20):
|
40 |
+
"""
|
41 |
+
Perform Reciprocal Rank Fusion to combine dense and sparse retrieval results
|
42 |
+
|
43 |
+
Args:
|
44 |
+
dense_results: List of (doc_id, score) from dense retrieval
|
45 |
+
sparse_results: List of (doc_id, score) from sparse retrieval
|
46 |
+
k: RRF parameter (default 20)
|
47 |
+
|
48 |
+
Returns:
|
49 |
+
List of (doc_id, rrf_score) sorted by RRF score
|
50 |
+
"""
|
51 |
+
rrf_scores = defaultdict(float)
|
52 |
+
|
53 |
+
# Add scores from dense retrieval
|
54 |
+
for rank, (doc_id, _) in enumerate(dense_results, 1):
|
55 |
+
rrf_scores[doc_id] += 1 / (k + rank)
|
56 |
+
|
57 |
+
# Add scores from sparse retrieval
|
58 |
+
for rank, (doc_id, _) in enumerate(sparse_results, 1):
|
59 |
+
rrf_scores[doc_id] += 1 / (k + rank)
|
60 |
+
|
61 |
+
# Sort by RRF score
|
62 |
+
sorted_results = sorted(rrf_scores.items(), key=lambda x: x[1], reverse=True)
|
63 |
+
return sorted_results
|
64 |
+
|
65 |
+
|
66 |
+
def normalized_score_fusion(dense_results, sparse_results, dense_weight=0.5, sparse_weight=0.5):
|
67 |
+
"""
|
68 |
+
Perform Normalized Score Fusion (NSF) with z-score normalization
|
69 |
+
|
70 |
+
Args:
|
71 |
+
dense_results: List of (doc_id, score) from dense retrieval
|
72 |
+
sparse_results: List of (doc_id, score) from sparse retrieval
|
73 |
+
dense_weight: Weight for dense scores (default 0.5)
|
74 |
+
sparse_weight: Weight for sparse scores (default 0.5)
|
75 |
+
|
76 |
+
Returns:
|
77 |
+
List of (doc_id, normalized_score) sorted by normalized score
|
78 |
+
"""
|
79 |
+
# Extract scores for normalization
|
80 |
+
dense_scores = np.array([score for _, score in dense_results])
|
81 |
+
sparse_scores = np.array([score for _, score in sparse_results])
|
82 |
+
|
83 |
+
# Z-score normalization (handle edge cases)
|
84 |
+
if len(dense_scores) > 1 and np.std(dense_scores) > 1e-10:
|
85 |
+
dense_scores_norm = stats.zscore(dense_scores)
|
86 |
+
else:
|
87 |
+
dense_scores_norm = np.zeros_like(dense_scores)
|
88 |
+
|
89 |
+
if len(sparse_scores) > 1 and np.std(sparse_scores) > 1e-10:
|
90 |
+
sparse_scores_norm = stats.zscore(sparse_scores)
|
91 |
+
else:
|
92 |
+
sparse_scores_norm = np.zeros_like(sparse_scores)
|
93 |
+
|
94 |
+
# Create dictionaries for normalized scores
|
95 |
+
dense_norm_dict = {doc_id: score for (doc_id, _), score in zip(dense_results, dense_scores_norm)}
|
96 |
+
sparse_norm_dict = {doc_id: score for (doc_id, _), score in zip(sparse_results, sparse_scores_norm)}
|
97 |
+
|
98 |
+
# Combine all unique document IDs
|
99 |
+
all_doc_ids = set()
|
100 |
+
all_doc_ids.update(doc_id for doc_id, _ in dense_results)
|
101 |
+
all_doc_ids.update(doc_id for doc_id, _ in sparse_results)
|
102 |
+
|
103 |
+
# Calculate weighted normalized scores
|
104 |
+
nsf_scores = {}
|
105 |
+
for doc_id in all_doc_ids:
|
106 |
+
dense_norm_score = dense_norm_dict.get(doc_id, 0.0)
|
107 |
+
sparse_norm_score = sparse_norm_dict.get(doc_id, 0.0)
|
108 |
+
|
109 |
+
# Weighted combination of normalized scores
|
110 |
+
nsf_scores[doc_id] = (dense_weight * dense_norm_score +
|
111 |
+
sparse_weight * sparse_norm_score)
|
112 |
+
|
113 |
+
# Sort by NSF score
|
114 |
+
sorted_results = sorted(nsf_scores.items(), key=lambda x: x[1], reverse=True)
|
115 |
+
return sorted_results
|
116 |
+
|
117 |
+
|
118 |
+
# Search part
|
119 |
+
## Currentl to try to speed up things, I keep only non-zero values for the sparse search
|
120 |
+
## Need to optimise the sparse research part (about 25-30s currently...)
|
121 |
+
## Tom must surely have some tips on this point
|
122 |
+
sparse_index = {}
|
123 |
+
for i, sparse_emd in enumerate(wikipedia_dataset["sparse_emb"]):
|
124 |
+
# Convert sparse_emd to a NumPy array
|
125 |
+
sparse_emd = np.array(sparse_emd)
|
126 |
+
|
127 |
+
# Only store non-zero values and their indices
|
128 |
+
non_zero_indices = np.nonzero(sparse_emd)[0]
|
129 |
+
non_zero_values = sparse_emd[non_zero_indices]
|
130 |
+
sparse_index[i] = (non_zero_indices, non_zero_values)
|
131 |
+
|
132 |
+
|
133 |
+
def sparse_search(query, top_k=20):
|
134 |
+
"""
|
135 |
+
Perform sparse retrieval using SPLADE representations with a dictionary-based sparse index.
|
136 |
+
"""
|
137 |
+
# Encode the query, the output is a sparse torch.Tensor
|
138 |
+
query_sparse_vector = sparse_model.encode(query, convert_to_numpy=False)
|
139 |
+
|
140 |
+
# Convert the sparse tensor to a dense format before using np.nonzero
|
141 |
+
query_sparse_vector = query_sparse_vector.to_dense().numpy()
|
142 |
+
|
143 |
+
# Get non-zero values and indices for the query
|
144 |
+
query_non_zero_indices = np.nonzero(query_sparse_vector)[0]
|
145 |
+
query_non_zero_values = query_sparse_vector[query_non_zero_indices]
|
146 |
+
|
147 |
+
# Compute dot product similarity efficiently
|
148 |
+
scores = defaultdict(float)
|
149 |
+
for doc_id, (doc_indices, doc_values) in sparse_index.items():
|
150 |
+
# Find the intersection of non-zero indices
|
151 |
+
common_indices = np.intersect1d(query_non_zero_indices, doc_indices)
|
152 |
+
# Compute dot product for common indices only
|
153 |
+
for idx in common_indices:
|
154 |
+
query_val = query_non_zero_values[np.where(query_non_zero_indices == idx)[0][0]]
|
155 |
+
doc_val = doc_values[np.where(doc_indices == idx)[0][0]]
|
156 |
+
scores[doc_id] += query_val * doc_val
|
157 |
+
|
158 |
+
# Sort and get top_k
|
159 |
+
sorted_scores = sorted(scores.items(), key=lambda x: x[1], reverse=True)
|
160 |
+
return sorted_scores[:top_k]
|
161 |
|
162 |
|
163 |
+
# Final search combining dense and sparse results
|
164 |
+
def search(query, top_k: int = 20, rescore_multiplier: int = 1, use_approx: bool = False,
|
165 |
+
fusion_method: str = "rrf", rrf_k: int = 20, dense_weight: float = 0.5, sparse_weight: float = 0.5):
|
166 |
+
total_start_time = time.time()
|
167 |
+
|
168 |
+
# 1. Dense retrieval pipeline (existing code)
|
169 |
start_time = time.time()
|
170 |
+
query_embedding = dense_model.encode(query, prompt="query: ")
|
171 |
embed_time = time.time() - start_time
|
172 |
|
|
|
173 |
start_time = time.time()
|
174 |
query_embedding_ubinary = quantize_embeddings(query_embedding.reshape(1, -1), "ubinary")
|
175 |
quantize_time = time.time() - start_time
|
176 |
|
|
|
177 |
index = binary_ivf if use_approx else binary_index
|
178 |
start_time = time.time()
|
179 |
+
_scores, binary_ids = index.search(query_embedding_ubinary, top_k * rescore_multiplier * 2) # Get more for fusion
|
180 |
binary_ids = binary_ids[0]
|
181 |
+
dense_search_time = time.time() - start_time
|
182 |
|
|
|
183 |
start_time = time.time()
|
184 |
int8_embeddings = int8_view[binary_ids].astype(int)
|
185 |
load_time = time.time() - start_time
|
186 |
|
|
|
187 |
start_time = time.time()
|
188 |
scores = query_embedding @ int8_embeddings.T
|
189 |
rescore_time = time.time() - start_time
|
190 |
|
191 |
+
# Prepare dense results
|
192 |
+
dense_results = [(binary_ids[i], scores[i]) for i in range(len(binary_ids))]
|
193 |
+
dense_results.sort(key=lambda x: x[1], reverse=True)
|
194 |
+
|
195 |
+
timing_info = {
|
196 |
+
"Temps pour créer l'embedding de la requête (dense)": f"{embed_time:.4f} s",
|
197 |
+
"Temps pour la quantification": f"{quantize_time:.4f} s",
|
198 |
+
"Temps pour effectuer la recherche dense": f"{dense_search_time:.4f} s",
|
199 |
+
"Temps de chargement": f"{load_time:.4f} s",
|
200 |
+
"Temps de rescorage": f"{rescore_time:.4f} s",
|
201 |
+
}
|
202 |
+
|
203 |
+
|
204 |
+
if fusion_method != "dense_only":
|
205 |
+
# 2. Sparse retrieval pipeline
|
206 |
+
start_time = time.time()
|
207 |
+
sparse_results = sparse_search(query, top_k * rescore_multiplier * 2)
|
208 |
+
sparse_search_time = time.time() - start_time
|
209 |
+
|
210 |
+
# 3. Apply selected fusion method
|
211 |
+
start_time = time.time()
|
212 |
+
if fusion_method == "rrf":
|
213 |
+
fusion_results = reciprocal_rank_fusion(dense_results, sparse_results, k=rrf_k)
|
214 |
+
fusion_method_name = f"RRF (k={rrf_k})"
|
215 |
+
elif fusion_method == "nsf":
|
216 |
+
fusion_results = normalized_score_fusion(dense_results, sparse_results,
|
217 |
+
dense_weight=dense_weight, sparse_weight=sparse_weight)
|
218 |
+
fusion_method_name = f"NSF Z-score (dense={dense_weight:.1f}, sparse={sparse_weight:.1f})"
|
219 |
+
else:
|
220 |
+
fusion_results = dense_results # fallback
|
221 |
+
fusion_method_name = "Dense uniquement (fallback)"
|
222 |
+
|
223 |
+
fusion_time = time.time() - start_time
|
224 |
+
|
225 |
+
# Use fusion results
|
226 |
+
final_results = fusion_results[:top_k * rescore_multiplier]
|
227 |
+
final_doc_ids = [doc_id for doc_id, _ in final_results]
|
228 |
+
final_scores = [score for _, score in final_results]
|
229 |
+
|
230 |
+
timing_info.update({
|
231 |
+
# Add time to create query embedding with sparse model
|
232 |
+
# Do quantification for sparse model as dense one?
|
233 |
+
"Temps pour la recherche sparse": f"{sparse_search_time:.4f} s",
|
234 |
+
"Temps pour la fusion": f"{fusion_time:.4f} s",
|
235 |
+
})
|
236 |
+
timing_info["Méthode de fusion utilisée"] = fusion_method_name
|
237 |
+
else:
|
238 |
+
# Use only dense results
|
239 |
+
final_doc_ids = [doc_id for doc_id, _ in dense_results[:top_k * rescore_multiplier]]
|
240 |
+
final_scores = [score for _, score in dense_results[:top_k * rescore_multiplier]]
|
241 |
+
timing_info["Méthode de fusion utilisée"] = "Dense uniquement"
|
242 |
+
|
243 |
+
# 4. Prepare final results
|
244 |
start_time = time.time()
|
245 |
+
try:
|
246 |
+
top_k_titles, top_k_texts = zip(*[(wikipedia_dataset[int(idx)]["title"], wikipedia_dataset[int(idx)]["text"])
|
247 |
+
for idx in final_doc_ids[:top_k]])
|
248 |
+
|
249 |
+
# Create DataFrame with results
|
250 |
+
df = pd.DataFrame({
|
251 |
+
"Score_paragraphe": [round(float(score), 4) for score in final_scores[:top_k]],
|
252 |
+
"Titre": top_k_titles,
|
253 |
+
"Texte": top_k_texts
|
254 |
+
})
|
255 |
+
|
256 |
+
# Calculate article scores (sum of paragraph scores)
|
257 |
+
score_sum = df.groupby('Titre')['Score_paragraphe'].sum().reset_index()
|
258 |
+
df = pd.merge(df, score_sum, on='Titre', how='left')
|
259 |
+
df.rename(columns={'Score_paragraphe_y': 'Score_article', 'Score_paragraphe_x': 'Score_paragraphe'}, inplace=True)
|
260 |
+
df = df[["Score_article", "Score_paragraphe", "Titre", "Texte"]]
|
261 |
+
df = df.sort_values('Score_article', ascending=False)
|
262 |
+
|
263 |
+
except Exception as e:
|
264 |
+
print(f"Error creating results DataFrame: {e}")
|
265 |
+
df = pd.DataFrame({"Error": [f"No results found or error processing results: {e}"]})
|
266 |
+
|
267 |
sort_time = time.time() - start_time
|
268 |
+
total_time = time.time() - total_start_time
|
269 |
+
|
270 |
+
timing_info.update({
|
271 |
+
"Temps pour afficher les résultats": f"{sort_time:.4f} s",
|
272 |
+
"Temps total": f"{total_time:.4f} s",
|
273 |
+
})
|
274 |
|
275 |
+
return df, timing_info
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
276 |
|
277 |
|
278 |
+
with gr.Blocks(title="Requêter Wikipedia avec Fusion Hybride 🔍") as demo:
|
279 |
|
280 |
gr.Markdown(
|
281 |
"""
|
282 |
## Requêter Wikipedia en temps réel 🔍
|
|
|
283 |
Ce démonstrateur permet de requêter un corpus composé des 250K paragraphes les plus consultés du Wikipédia francophone.
|
284 |
Les résultats sont renvoyés en temps réel via un pipeline tournant sur un CPU 🚀
|
285 |
Nous nous sommes grandement inspirés du Space [quantized-retrieval](https://huggingface.co/spaces/sentence-transformers/quantized-retrieval) conçu par [Tom Aarsen](https://huggingface.co/tomaarsen) 🤗
|
|
|
297 |
</details>
|
298 |
|
299 |
<details><summary>2. Détails le pipeline</summary>
|
300 |
+
A écrire quand ça sera terminé.
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
301 |
</details>
|
302 |
"""
|
303 |
)
|
304 |
+
|
305 |
with gr.Row():
|
306 |
with gr.Column(scale=75):
|
307 |
query = gr.Textbox(
|
|
|
312 |
use_approx = gr.Radio(
|
313 |
choices=[("Exacte", False), ("Approximative", True)],
|
314 |
value=True,
|
315 |
+
label="Type de recherche dense",
|
316 |
)
|
317 |
|
318 |
with gr.Row():
|
|
|
322 |
maximum=40,
|
323 |
step=1,
|
324 |
value=15,
|
325 |
+
label="Nombre de documents à retourner",
|
|
|
326 |
)
|
327 |
with gr.Column(scale=2):
|
328 |
rescore_multiplier = gr.Slider(
|
|
|
331 |
step=1,
|
332 |
value=1,
|
333 |
label="Coefficient de rescorage",
|
334 |
+
info="Augmente le nombre de candidats avant fusion",
|
335 |
+
)
|
336 |
+
|
337 |
+
with gr.Row():
|
338 |
+
with gr.Column(scale=3):
|
339 |
+
fusion_method = gr.Radio(
|
340 |
+
choices=[
|
341 |
+
("Dense uniquement", "dense_only"),
|
342 |
+
("Fusion RRF", "rrf"),
|
343 |
+
("Fusion NSF Z-score", "nsf")
|
344 |
+
],
|
345 |
+
value="rrf",
|
346 |
+
label="Méthode de fusion",
|
347 |
+
info="Choisissez comment combiner les résultats des modèles dense et sparse"
|
348 |
)
|
349 |
+
with gr.Column(scale=2):
|
350 |
+
rrf_k = gr.Slider(
|
351 |
+
minimum=1,
|
352 |
+
maximum=200,
|
353 |
+
step=1,
|
354 |
+
value=60,
|
355 |
+
label="Paramètre k pour RRF",
|
356 |
+
info="Plus k est élevé, moins les rangs ont d'importance",
|
357 |
+
visible=True
|
358 |
+
)
|
359 |
+
|
360 |
+
with gr.Row():
|
361 |
+
with gr.Column(scale=2):
|
362 |
+
dense_weight = gr.Slider(
|
363 |
+
minimum=0.0,
|
364 |
+
maximum=1.0,
|
365 |
+
step=0.1,
|
366 |
+
value=0.5,
|
367 |
+
label="Poids Dense (NSF)",
|
368 |
+
info="Importance des résultats du modèle dense",
|
369 |
+
visible=False
|
370 |
+
)
|
371 |
+
with gr.Column(scale=2):
|
372 |
+
sparse_weight = gr.Slider(
|
373 |
+
minimum=0.0,
|
374 |
+
maximum=1.0,
|
375 |
+
step=0.1,
|
376 |
+
value=0.5,
|
377 |
+
label="Poids Sparse (NSF)",
|
378 |
+
info="Importance des résultats du modèle sparse",
|
379 |
+
visible=False
|
380 |
+
)
|
381 |
+
|
382 |
+
# JavaScript to show/hide parameters based on fusion method
|
383 |
+
fusion_method.change(
|
384 |
+
fn=lambda method: (
|
385 |
+
gr.update(visible=(method == "rrf")), # rrf_k
|
386 |
+
gr.update(visible=(method == "nsf")), # dense_weight
|
387 |
+
gr.update(visible=(method == "nsf")) # sparse_weight
|
388 |
+
),
|
389 |
+
inputs=[fusion_method],
|
390 |
+
outputs=[rrf_k, dense_weight, sparse_weight]
|
391 |
+
)
|
392 |
|
393 |
+
search_button = gr.Button(value="Rechercher", variant="primary")
|
394 |
|
395 |
output = gr.Dataframe(headers=["Score_article", "Score_paragraphe", "Titre", "Texte"], datatype="markdown")
|
396 |
+
json = gr.JSON(label="Informations de performance")
|
|
|
|
|
|
|
|
|
397 |
|
398 |
+
query.submit(search, inputs=[query, top_k, rescore_multiplier, use_approx, fusion_method, rrf_k, dense_weight, sparse_weight], outputs=[output, json])
|
399 |
+
search_button.click(search, inputs=[query, top_k, rescore_multiplier, use_approx, fusion_method, rrf_k, dense_weight, sparse_weight], outputs=[output, json])
|
400 |
+
|
401 |
demo.queue()
|
402 |
+
demo.launch()
|