Update app.py
Browse files
app.py
CHANGED
@@ -5,38 +5,35 @@ import docx
|
|
5 |
import nltk
|
6 |
import gradio as gr
|
7 |
from langchain_huggingface import HuggingFaceEmbeddings
|
8 |
-
from langchain_community.embeddings import
|
9 |
-
OpenAIEmbeddings,
|
10 |
-
CohereEmbeddings,
|
11 |
-
)
|
12 |
from langchain_openai import OpenAIEmbeddings
|
13 |
from langchain_community.vectorstores import FAISS, Chroma
|
14 |
-
from langchain_text_splitters import
|
15 |
-
RecursiveCharacterTextSplitter,
|
16 |
-
TokenTextSplitter,
|
17 |
-
)
|
18 |
from typing import List, Dict, Any
|
19 |
import pandas as pd
|
|
|
20 |
import re
|
21 |
from nltk.corpus import stopwords
|
22 |
from nltk.tokenize import word_tokenize
|
23 |
from nltk.stem import SnowballStemmer
|
24 |
-
import jellyfish
|
25 |
from gensim.models import Word2Vec
|
26 |
from gensim.models.fasttext import FastText
|
27 |
from collections import Counter
|
28 |
from tokenizers import Tokenizer
|
29 |
-
from tokenizers.models import
|
30 |
-
from tokenizers.trainers import
|
31 |
-
|
32 |
-
|
|
|
|
|
|
|
|
|
|
|
33 |
|
|
|
34 |
def download_nltk_resources():
|
35 |
-
resources = [
|
36 |
-
'punkt',
|
37 |
-
'stopwords',
|
38 |
-
'snowball_data',
|
39 |
-
]
|
40 |
for resource in resources:
|
41 |
try:
|
42 |
nltk.download(resource, quiet=True)
|
@@ -45,47 +42,87 @@ def download_nltk_resources():
|
|
45 |
|
46 |
download_nltk_resources()
|
47 |
|
|
|
48 |
|
49 |
-
|
50 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
51 |
|
52 |
-
|
|
|
|
|
|
|
53 |
|
54 |
-
|
55 |
-
|
56 |
-
|
57 |
-
|
58 |
-
|
59 |
-
|
60 |
-
|
61 |
-
|
62 |
-
|
63 |
-
|
64 |
-
|
65 |
-
'Cohere': {
|
66 |
-
'embed-multilingual-v2.0': "embed-multilingual-v2.0"
|
67 |
-
}
|
68 |
-
}
|
69 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
70 |
def simple_tokenize(text):
|
71 |
-
"""Simple tokenization fallback method."""
|
72 |
return text.split()
|
73 |
|
74 |
def preprocess_text(text, lang='german'):
|
75 |
-
# Convert to lowercase
|
76 |
text = text.lower()
|
77 |
-
|
78 |
-
# Remove special characters and digits
|
79 |
text = re.sub(r'[^a-zA-Z\s]', '', text)
|
80 |
|
81 |
-
# Tokenize
|
82 |
try:
|
83 |
tokens = word_tokenize(text, language=lang)
|
84 |
except LookupError:
|
85 |
print(f"Warning: NLTK punkt tokenizer for {lang} not found. Using simple tokenization.")
|
86 |
tokens = simple_tokenize(text)
|
87 |
|
88 |
-
# Remove stopwords
|
89 |
try:
|
90 |
stop_words = set(stopwords.words(lang))
|
91 |
except LookupError:
|
@@ -93,7 +130,6 @@ def preprocess_text(text, lang='german'):
|
|
93 |
stop_words = set()
|
94 |
tokens = [token for token in tokens if token not in stop_words]
|
95 |
|
96 |
-
# Stemming
|
97 |
try:
|
98 |
stemmer = SnowballStemmer(lang)
|
99 |
tokens = [stemmer.stem(token) for token in tokens]
|
@@ -107,44 +143,34 @@ def phonetic_match(text, query, method='koelner_phonetik'):
|
|
107 |
text_phonetic = jellyfish.cologne_phonetic(text)
|
108 |
query_phonetic = jellyfish.cologne_phonetic(query)
|
109 |
return jellyfish.jaro_winkler(text_phonetic, query_phonetic)
|
110 |
-
# Add other phonetic methods as needed
|
111 |
return 0
|
112 |
|
113 |
-
|
114 |
-
|
115 |
-
|
116 |
-
|
117 |
-
if ext == '.pdf':
|
118 |
-
return FileHandler._extract_from_pdf(file_path)
|
119 |
-
elif ext == '.docx':
|
120 |
-
return FileHandler._extract_from_docx(file_path)
|
121 |
-
elif ext == '.txt':
|
122 |
-
return FileHandler._extract_from_txt(file_path)
|
123 |
-
else:
|
124 |
-
raise ValueError(f"Unsupported file type: {ext}")
|
125 |
|
126 |
-
|
127 |
-
|
128 |
-
with pdfplumber.open(file_path) as pdf:
|
129 |
-
return ' '.join([page.extract_text() for page in pdf.pages])
|
130 |
|
131 |
-
|
132 |
-
|
133 |
-
doc = docx.Document(file_path)
|
134 |
-
return ' '.join([para.text for para in doc.paragraphs])
|
135 |
|
136 |
-
|
137 |
-
def _extract_from_txt(file_path):
|
138 |
-
with open(file_path, 'r', encoding='utf-8') as f:
|
139 |
-
return f.read()
|
140 |
|
|
|
|
|
|
|
|
|
|
|
141 |
def get_embedding_model(model_type, model_name):
|
|
|
142 |
if model_type == 'HuggingFace':
|
143 |
-
return HuggingFaceEmbeddings(model_name=
|
144 |
elif model_type == 'OpenAI':
|
145 |
-
return OpenAIEmbeddings(model=
|
146 |
elif model_type == 'Cohere':
|
147 |
-
return CohereEmbeddings(model=
|
148 |
else:
|
149 |
raise ValueError(f"Unsupported model type: {model_type}")
|
150 |
|
@@ -160,6 +186,7 @@ def get_text_splitter(split_strategy, chunk_size, overlap_size, custom_separator
|
|
160 |
else:
|
161 |
raise ValueError(f"Unsupported split strategy: {split_strategy}")
|
162 |
|
|
|
163 |
def get_vector_store(vector_store_type, chunks, embedding_model):
|
164 |
if vector_store_type == 'FAISS':
|
165 |
return FAISS.from_texts(chunks, embedding_model)
|
@@ -179,7 +206,8 @@ def get_retriever(vector_store, search_type, search_kwargs):
|
|
179 |
else:
|
180 |
raise ValueError(f"Unsupported search type: {search_type}")
|
181 |
|
182 |
-
|
|
|
183 |
if file_path:
|
184 |
text = FileHandler.extract_text(file_path)
|
185 |
else:
|
@@ -188,8 +216,11 @@ def process_files(file_path, model_type, model_name, split_strategy, chunk_size,
|
|
188 |
file_path = os.path.join(FILES_DIR, file)
|
189 |
text += FileHandler.extract_text(file_path)
|
190 |
|
191 |
-
|
192 |
-
|
|
|
|
|
|
|
193 |
|
194 |
text_splitter = get_text_splitter(split_strategy, chunk_size, overlap_size, custom_separators)
|
195 |
chunks = text_splitter.split_text(text)
|
@@ -199,7 +230,6 @@ def process_files(file_path, model_type, model_name, split_strategy, chunk_size,
|
|
199 |
return chunks, embedding_model, len(text.split())
|
200 |
|
201 |
def search_embeddings(chunks, embedding_model, vector_store_type, search_type, query, top_k, lang='german', phonetic_weight=0.3):
|
202 |
-
# Preprocess the query
|
203 |
preprocessed_query = preprocess_text(query, lang)
|
204 |
|
205 |
vector_store = get_vector_store(vector_store_type, chunks, embedding_model)
|
@@ -208,15 +238,18 @@ def search_embeddings(chunks, embedding_model, vector_store_type, search_type, q
|
|
208 |
start_time = time.time()
|
209 |
results = retriever.invoke(preprocessed_query)
|
210 |
|
211 |
-
|
212 |
-
|
213 |
-
|
214 |
-
|
|
|
|
|
215 |
|
216 |
end_time = time.time()
|
217 |
|
218 |
return results[:top_k], end_time - start_time, vector_store
|
219 |
|
|
|
220 |
def calculate_statistics(results, search_time, vector_store, num_tokens, embedding_model, query, top_k):
|
221 |
stats = {
|
222 |
"num_results": len(results),
|
@@ -230,64 +263,58 @@ def calculate_statistics(results, search_time, vector_store, num_tokens, embeddi
|
|
230 |
"top_k": top_k,
|
231 |
}
|
232 |
|
233 |
-
# Calculate diversity of results
|
234 |
if len(results) > 1:
|
235 |
embeddings = [embedding_model.embed_query(doc.page_content) for doc in results]
|
236 |
-
pairwise_similarities =
|
237 |
stats["result_diversity"] = 1 - np.mean(pairwise_similarities[np.triu_indices(len(embeddings), k=1)])
|
|
|
|
|
|
|
|
|
|
|
|
|
238 |
else:
|
239 |
stats["result_diversity"] = "N/A"
|
|
|
240 |
|
241 |
-
# Calculate rank correlation between embedding similarity and result order
|
242 |
query_embedding = embedding_model.embed_query(query)
|
243 |
result_embeddings = [embedding_model.embed_query(doc.page_content) for doc in results]
|
244 |
-
similarities = [
|
245 |
rank_correlation, _ = spearmanr(similarities, range(len(similarities)))
|
246 |
stats["rank_correlation"] = rank_correlation
|
247 |
|
248 |
return stats
|
249 |
|
250 |
-
|
251 |
-
|
252 |
-
|
253 |
-
|
254 |
-
if model_type == 'word2vec':
|
255 |
-
model = Word2Vec(sentences=tokenized_texts, vector_size=vector_size, window=window, min_count=min_count, workers=4)
|
256 |
-
elif model_type == 'fasttext':
|
257 |
-
model = FastText(sentences=tokenized_texts, vector_size=vector_size, window=window, min_count=min_count, workers=4)
|
258 |
-
else:
|
259 |
-
raise ValueError("Unsupported model type")
|
260 |
-
|
261 |
-
return model
|
262 |
-
|
263 |
-
class CustomEmbeddings(HuggingFaceEmbeddings):
|
264 |
-
def __init__(self, model_path):
|
265 |
-
self.model = Word2Vec.load(model_path) # or FastText.load() for FastText models
|
266 |
|
267 |
-
|
268 |
-
|
|
|
269 |
|
270 |
-
|
271 |
-
|
272 |
-
|
273 |
-
def optimize_vocabulary(texts, vocab_size=10000, min_frequency=2):
|
274 |
-
# Count word frequencies
|
275 |
-
word_freq = Counter(word for text in texts for word in text.split())
|
276 |
|
277 |
-
|
278 |
-
|
279 |
-
|
280 |
-
for text in texts
|
281 |
-
]
|
282 |
|
283 |
-
|
284 |
-
|
285 |
-
|
286 |
-
|
|
|
|
|
|
|
|
|
|
|
287 |
|
288 |
-
|
|
|
289 |
|
290 |
-
|
|
|
291 |
all_results = []
|
292 |
all_stats = []
|
293 |
settings = {
|
@@ -313,7 +340,8 @@ def compare_embeddings(file, query, model_types, model_names, split_strategy, ch
|
|
313 |
chunk_size,
|
314 |
overlap_size,
|
315 |
custom_separators.split(',') if custom_separators else None,
|
316 |
-
lang
|
|
|
317 |
)
|
318 |
|
319 |
if use_custom_embedding:
|
@@ -324,7 +352,6 @@ def compare_embeddings(file, query, model_types, model_names, split_strategy, ch
|
|
324 |
tokenizer, optimized_chunks = optimize_vocabulary(chunks)
|
325 |
chunks = optimized_chunks
|
326 |
|
327 |
-
|
328 |
results, search_time, vector_store = search_embeddings(
|
329 |
chunks,
|
330 |
embedding_model,
|
@@ -347,7 +374,10 @@ def compare_embeddings(file, query, model_types, model_names, split_strategy, ch
|
|
347 |
results_df = pd.DataFrame(all_results)
|
348 |
stats_df = pd.DataFrame(all_stats)
|
349 |
|
350 |
-
|
|
|
|
|
|
|
351 |
|
352 |
def format_results(results, stats):
|
353 |
formatted_results = []
|
@@ -355,53 +385,22 @@ def format_results(results, stats):
|
|
355 |
result = {
|
356 |
"Model": stats["model"],
|
357 |
"Content": doc.page_content,
|
|
|
358 |
**doc.metadata,
|
359 |
**{k: v for k, v in stats.items() if k not in ["model"]}
|
360 |
}
|
361 |
formatted_results.append(result)
|
362 |
return formatted_results
|
363 |
|
364 |
-
|
365 |
-
import seaborn as sns
|
366 |
-
from sklearn.manifold import TSNE
|
367 |
-
|
368 |
-
def visualize_results(results_df, stats_df):
|
369 |
-
# Create a figure with subplots
|
370 |
-
fig, axs = plt.subplots(2, 2, figsize=(20, 20))
|
371 |
-
|
372 |
-
# 1. Bar plot of search times
|
373 |
-
sns.barplot(x='model', y='search_time', data=stats_df, ax=axs[0, 0])
|
374 |
-
axs[0, 0].set_title('Search Time by Model')
|
375 |
-
axs[0, 0].set_xticklabels(axs[0, 0].get_xticklabels(), rotation=45, ha='right')
|
376 |
-
|
377 |
-
# 2. Scatter plot of result diversity vs. rank correlation
|
378 |
-
sns.scatterplot(x='result_diversity', y='rank_correlation', hue='model', data=stats_df, ax=axs[0, 1])
|
379 |
-
axs[0, 1].set_title('Result Diversity vs. Rank Correlation')
|
380 |
-
|
381 |
-
# 3. Box plot of content lengths
|
382 |
-
sns.boxplot(x='model', y='content_length', data=results_df, ax=axs[1, 0])
|
383 |
-
axs[1, 0].set_title('Distribution of Result Content Lengths')
|
384 |
-
axs[1, 0].set_xticklabels(axs[1, 0].get_xticklabels(), rotation=45, ha='right')
|
385 |
-
|
386 |
-
# 4. t-SNE visualization of embeddings
|
387 |
-
embeddings = np.array(results_df['embedding'].tolist())
|
388 |
-
tsne = TSNE(n_components=2, random_state=42)
|
389 |
-
embeddings_2d = tsne.fit_transform(embeddings)
|
390 |
-
|
391 |
-
sns.scatterplot(x=embeddings_2d[:, 0], y=embeddings_2d[:, 1], hue=results_df['model'], ax=axs[1, 1])
|
392 |
-
axs[1, 1].set_title('t-SNE Visualization of Result Embeddings')
|
393 |
-
|
394 |
-
plt.tight_layout()
|
395 |
-
return fig
|
396 |
-
|
397 |
def launch_interface(share=True):
|
398 |
iface = gr.Interface(
|
399 |
fn=compare_embeddings,
|
400 |
inputs=[
|
401 |
gr.File(label="Upload File (Optional)"),
|
402 |
gr.Textbox(label="Search Query"),
|
403 |
-
gr.CheckboxGroup(choices=list(
|
404 |
-
gr.CheckboxGroup(choices=[model for models in
|
405 |
gr.Radio(choices=["token", "recursive"], label="Split Strategy", value="recursive"),
|
406 |
gr.Slider(100, 1000, step=100, value=500, label="Chunk Size"),
|
407 |
gr.Slider(0, 100, step=10, value=50, label="Overlap Size"),
|
@@ -412,7 +411,8 @@ def launch_interface(share=True):
|
|
412 |
gr.Dropdown(choices=["german", "english", "french"], label="Language", value="german"),
|
413 |
gr.Checkbox(label="Use Custom Embedding", value=False),
|
414 |
gr.Checkbox(label="Optimize Vocabulary", value=False),
|
415 |
-
gr.Slider(0, 1, step=0.1, value=0.3, label="Phonetic Matching Weight")
|
|
|
416 |
],
|
417 |
outputs=[
|
418 |
gr.Dataframe(label="Results", interactive=False),
|
@@ -426,7 +426,21 @@ def launch_interface(share=True):
|
|
426 |
tutorial_md = """
|
427 |
# Advanced Embedding Comparison Tool Tutorial
|
428 |
|
429 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
430 |
"""
|
431 |
|
432 |
iface = gr.TabbedInterface(
|
|
|
5 |
import nltk
|
6 |
import gradio as gr
|
7 |
from langchain_huggingface import HuggingFaceEmbeddings
|
8 |
+
from langchain_community.embeddings import CohereEmbeddings
|
|
|
|
|
|
|
9 |
from langchain_openai import OpenAIEmbeddings
|
10 |
from langchain_community.vectorstores import FAISS, Chroma
|
11 |
+
from langchain_text_splitters import RecursiveCharacterTextSplitter, TokenTextSplitter
|
|
|
|
|
|
|
12 |
from typing import List, Dict, Any
|
13 |
import pandas as pd
|
14 |
+
import numpy as np
|
15 |
import re
|
16 |
from nltk.corpus import stopwords
|
17 |
from nltk.tokenize import word_tokenize
|
18 |
from nltk.stem import SnowballStemmer
|
19 |
+
import jellyfish
|
20 |
from gensim.models import Word2Vec
|
21 |
from gensim.models.fasttext import FastText
|
22 |
from collections import Counter
|
23 |
from tokenizers import Tokenizer
|
24 |
+
from tokenizers.models import WordLevel
|
25 |
+
from tokenizers.trainers import WordLevelTrainer
|
26 |
+
from tokenizers.pre_tokenizers import Whitespace
|
27 |
+
import matplotlib.pyplot as plt
|
28 |
+
import seaborn as sns
|
29 |
+
from sklearn.manifold import TSNE
|
30 |
+
from sklearn.metrics import silhouette_score
|
31 |
+
from scipy.stats import spearmanr
|
32 |
+
from functools import lru_cache
|
33 |
|
34 |
+
# NLTK Resource Download
|
35 |
def download_nltk_resources():
|
36 |
+
resources = ['punkt', 'stopwords', 'snowball_data']
|
|
|
|
|
|
|
|
|
37 |
for resource in resources:
|
38 |
try:
|
39 |
nltk.download(resource, quiet=True)
|
|
|
42 |
|
43 |
download_nltk_resources()
|
44 |
|
45 |
+
FILES_DIR = './files'
|
46 |
|
47 |
+
# Model Management
|
48 |
+
class ModelManager:
|
49 |
+
def __init__(self):
|
50 |
+
self.models = {
|
51 |
+
'HuggingFace': {
|
52 |
+
'e5-base-de': "danielheinz/e5-base-sts-en-de",
|
53 |
+
'paraphrase-miniLM': "paraphrase-multilingual-MiniLM-L12-v2",
|
54 |
+
'paraphrase-mpnet': "paraphrase-multilingual-mpnet-base-v2",
|
55 |
+
'gte-large': "gte-large",
|
56 |
+
'gbert-base': "gbert-base"
|
57 |
+
},
|
58 |
+
'OpenAI': {
|
59 |
+
'text-embedding-ada-002': "text-embedding-ada-002"
|
60 |
+
},
|
61 |
+
'Cohere': {
|
62 |
+
'embed-multilingual-v2.0': "embed-multilingual-v2.0"
|
63 |
+
}
|
64 |
+
}
|
65 |
|
66 |
+
def add_model(self, provider, name, model_path):
|
67 |
+
if provider not in self.models:
|
68 |
+
self.models[provider] = {}
|
69 |
+
self.models[provider][name] = model_path
|
70 |
|
71 |
+
def remove_model(self, provider, name):
|
72 |
+
if provider in self.models and name in self.models[provider]:
|
73 |
+
del self.models[provider][name]
|
74 |
+
|
75 |
+
def get_model(self, provider, name):
|
76 |
+
return self.models.get(provider, {}).get(name)
|
77 |
+
|
78 |
+
def list_models(self):
|
79 |
+
return {provider: list(models.keys()) for provider, models in self.models.items()}
|
80 |
+
|
81 |
+
model_manager = ModelManager()
|
|
|
|
|
|
|
|
|
82 |
|
83 |
+
# File Handling
|
84 |
+
class FileHandler:
|
85 |
+
@staticmethod
|
86 |
+
def extract_text(file_path):
|
87 |
+
ext = os.path.splitext(file_path)[-1].lower()
|
88 |
+
if ext == '.pdf':
|
89 |
+
return FileHandler._extract_from_pdf(file_path)
|
90 |
+
elif ext == '.docx':
|
91 |
+
return FileHandler._extract_from_docx(file_path)
|
92 |
+
elif ext == '.txt':
|
93 |
+
return FileHandler._extract_from_txt(file_path)
|
94 |
+
else:
|
95 |
+
raise ValueError(f"Unsupported file type: {ext}")
|
96 |
+
|
97 |
+
@staticmethod
|
98 |
+
def _extract_from_pdf(file_path):
|
99 |
+
with pdfplumber.open(file_path) as pdf:
|
100 |
+
return ' '.join([page.extract_text() for page in pdf.pages])
|
101 |
+
|
102 |
+
@staticmethod
|
103 |
+
def _extract_from_docx(file_path):
|
104 |
+
doc = docx.Document(file_path)
|
105 |
+
return ' '.join([para.text for para in doc.paragraphs])
|
106 |
+
|
107 |
+
@staticmethod
|
108 |
+
def _extract_from_txt(file_path):
|
109 |
+
with open(file_path, 'r', encoding='utf-8') as f:
|
110 |
+
return f.read()
|
111 |
+
|
112 |
+
# Text Processing
|
113 |
def simple_tokenize(text):
|
|
|
114 |
return text.split()
|
115 |
|
116 |
def preprocess_text(text, lang='german'):
|
|
|
117 |
text = text.lower()
|
|
|
|
|
118 |
text = re.sub(r'[^a-zA-Z\s]', '', text)
|
119 |
|
|
|
120 |
try:
|
121 |
tokens = word_tokenize(text, language=lang)
|
122 |
except LookupError:
|
123 |
print(f"Warning: NLTK punkt tokenizer for {lang} not found. Using simple tokenization.")
|
124 |
tokens = simple_tokenize(text)
|
125 |
|
|
|
126 |
try:
|
127 |
stop_words = set(stopwords.words(lang))
|
128 |
except LookupError:
|
|
|
130 |
stop_words = set()
|
131 |
tokens = [token for token in tokens if token not in stop_words]
|
132 |
|
|
|
133 |
try:
|
134 |
stemmer = SnowballStemmer(lang)
|
135 |
tokens = [stemmer.stem(token) for token in tokens]
|
|
|
143 |
text_phonetic = jellyfish.cologne_phonetic(text)
|
144 |
query_phonetic = jellyfish.cologne_phonetic(query)
|
145 |
return jellyfish.jaro_winkler(text_phonetic, query_phonetic)
|
|
|
146 |
return 0
|
147 |
|
148 |
+
# Custom Tokenizer
|
149 |
+
def create_custom_tokenizer(file_path):
|
150 |
+
with open(file_path, 'r', encoding='utf-8') as f:
|
151 |
+
text = f.read()
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
152 |
|
153 |
+
tokenizer = Tokenizer(WordLevel(unk_token="[UNK]"))
|
154 |
+
tokenizer.pre_tokenizer = Whitespace()
|
|
|
|
|
155 |
|
156 |
+
trainer = WordLevelTrainer(special_tokens=["[UNK]", "[CLS]", "[SEP]", "[PAD]", "[MASK]"])
|
157 |
+
tokenizer.train_from_iterator([text], trainer)
|
|
|
|
|
158 |
|
159 |
+
return tokenizer
|
|
|
|
|
|
|
160 |
|
161 |
+
def custom_tokenize(text, tokenizer):
|
162 |
+
return tokenizer.encode(text).tokens
|
163 |
+
|
164 |
+
# Embedding and Vector Store
|
165 |
+
@lru_cache(maxsize=None)
|
166 |
def get_embedding_model(model_type, model_name):
|
167 |
+
model_path = model_manager.get_model(model_type, model_name)
|
168 |
if model_type == 'HuggingFace':
|
169 |
+
return HuggingFaceEmbeddings(model_name=model_path)
|
170 |
elif model_type == 'OpenAI':
|
171 |
+
return OpenAIEmbeddings(model=model_path)
|
172 |
elif model_type == 'Cohere':
|
173 |
+
return CohereEmbeddings(model=model_path)
|
174 |
else:
|
175 |
raise ValueError(f"Unsupported model type: {model_type}")
|
176 |
|
|
|
186 |
else:
|
187 |
raise ValueError(f"Unsupported split strategy: {split_strategy}")
|
188 |
|
189 |
+
@lru_cache(maxsize=None)
|
190 |
def get_vector_store(vector_store_type, chunks, embedding_model):
|
191 |
if vector_store_type == 'FAISS':
|
192 |
return FAISS.from_texts(chunks, embedding_model)
|
|
|
206 |
else:
|
207 |
raise ValueError(f"Unsupported search type: {search_type}")
|
208 |
|
209 |
+
# Main Processing Functions
|
210 |
+
def process_files(file_path, model_type, model_name, split_strategy, chunk_size, overlap_size, custom_separators, lang='german', custom_tokenizer_file=None):
|
211 |
if file_path:
|
212 |
text = FileHandler.extract_text(file_path)
|
213 |
else:
|
|
|
216 |
file_path = os.path.join(FILES_DIR, file)
|
217 |
text += FileHandler.extract_text(file_path)
|
218 |
|
219 |
+
if custom_tokenizer_file:
|
220 |
+
tokenizer = create_custom_tokenizer(custom_tokenizer_file)
|
221 |
+
text = ' '.join(custom_tokenize(text, tokenizer))
|
222 |
+
else:
|
223 |
+
text = preprocess_text(text, lang)
|
224 |
|
225 |
text_splitter = get_text_splitter(split_strategy, chunk_size, overlap_size, custom_separators)
|
226 |
chunks = text_splitter.split_text(text)
|
|
|
230 |
return chunks, embedding_model, len(text.split())
|
231 |
|
232 |
def search_embeddings(chunks, embedding_model, vector_store_type, search_type, query, top_k, lang='german', phonetic_weight=0.3):
|
|
|
233 |
preprocessed_query = preprocess_text(query, lang)
|
234 |
|
235 |
vector_store = get_vector_store(vector_store_type, chunks, embedding_model)
|
|
|
238 |
start_time = time.time()
|
239 |
results = retriever.invoke(preprocessed_query)
|
240 |
|
241 |
+
def score_result(doc):
|
242 |
+
similarity_score = vector_store.similarity_search_with_score(doc.page_content, k=1)[0][1]
|
243 |
+
phonetic_score = phonetic_match(doc.page_content, query)
|
244 |
+
return (1 - phonetic_weight) * similarity_score + phonetic_weight * phonetic_score
|
245 |
+
|
246 |
+
results = sorted(results, key=score_result, reverse=True)
|
247 |
|
248 |
end_time = time.time()
|
249 |
|
250 |
return results[:top_k], end_time - start_time, vector_store
|
251 |
|
252 |
+
# Evaluation Metrics
|
253 |
def calculate_statistics(results, search_time, vector_store, num_tokens, embedding_model, query, top_k):
|
254 |
stats = {
|
255 |
"num_results": len(results),
|
|
|
263 |
"top_k": top_k,
|
264 |
}
|
265 |
|
|
|
266 |
if len(results) > 1:
|
267 |
embeddings = [embedding_model.embed_query(doc.page_content) for doc in results]
|
268 |
+
pairwise_similarities = np.inner(embeddings, embeddings)
|
269 |
stats["result_diversity"] = 1 - np.mean(pairwise_similarities[np.triu_indices(len(embeddings), k=1)])
|
270 |
+
|
271 |
+
# Silhouette Score
|
272 |
+
if len(embeddings) > 2:
|
273 |
+
stats["silhouette_score"] = silhouette_score(embeddings, range(len(embeddings)))
|
274 |
+
else:
|
275 |
+
stats["silhouette_score"] = "N/A"
|
276 |
else:
|
277 |
stats["result_diversity"] = "N/A"
|
278 |
+
stats["silhouette_score"] = "N/A"
|
279 |
|
|
|
280 |
query_embedding = embedding_model.embed_query(query)
|
281 |
result_embeddings = [embedding_model.embed_query(doc.page_content) for doc in results]
|
282 |
+
similarities = [np.inner(query_embedding, emb)[0] for emb in result_embeddings]
|
283 |
rank_correlation, _ = spearmanr(similarities, range(len(similarities)))
|
284 |
stats["rank_correlation"] = rank_correlation
|
285 |
|
286 |
return stats
|
287 |
|
288 |
+
# Visualization
|
289 |
+
def visualize_results(results_df, stats_df):
|
290 |
+
fig, axs = plt.subplots(2, 2, figsize=(20, 20))
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
291 |
|
292 |
+
sns.barplot(x='model', y='search_time', data=stats_df, ax=axs[0, 0])
|
293 |
+
axs[0, 0].set_title('Search Time by Model')
|
294 |
+
axs[0, 0].set_xticklabels(axs[0, 0].get_xticklabels(), rotation=45, ha='right')
|
295 |
|
296 |
+
sns.scatterplot(x='result_diversity', y='rank_correlation', hue='model', data=stats_df, ax=axs[0, 1])
|
297 |
+
axs[0, 1].set_title('Result Diversity vs. Rank Correlation')
|
|
|
|
|
|
|
|
|
298 |
|
299 |
+
sns.boxplot(x='model', y='avg_content_length', data=stats_df, ax=axs[1, 0])
|
300 |
+
axs[1, 0].set_title('Distribution of Result Content Lengths')
|
301 |
+
axs[1, 0].set_xticklabels(axs[1, 0].get_xticklabels(), rotation=45, ha='right')
|
|
|
|
|
302 |
|
303 |
+
embeddings = np.array([embedding for embedding in results_df['embedding'] if isinstance(embedding, np.ndarray)])
|
304 |
+
if len(embeddings) > 1:
|
305 |
+
tsne = TSNE(n_components=2, random_state=42)
|
306 |
+
embeddings_2d = tsne.fit_transform(embeddings)
|
307 |
+
|
308 |
+
sns.scatterplot(x=embeddings_2d[:, 0], y=embeddings_2d[:, 1], hue=results_df['model'][:len(embeddings)], ax=axs[1, 1])
|
309 |
+
axs[1, 1].set_title('t-SNE Visualization of Result Embeddings')
|
310 |
+
else:
|
311 |
+
axs[1, 1].text(0.5, 0.5, "Not enough data for t-SNE visualization", ha='center', va='center')
|
312 |
|
313 |
+
plt.tight_layout()
|
314 |
+
return fig
|
315 |
|
316 |
+
# Main Comparison Function
|
317 |
+
def compare_embeddings(file, query, model_types, model_names, split_strategy, chunk_size, overlap_size, custom_separators, vector_store_type, search_type, top_k, lang='german', use_custom_embedding=False, optimize_vocab=False, phonetic_weight=0.3, custom_tokenizer_file=None):
|
318 |
all_results = []
|
319 |
all_stats = []
|
320 |
settings = {
|
|
|
340 |
chunk_size,
|
341 |
overlap_size,
|
342 |
custom_separators.split(',') if custom_separators else None,
|
343 |
+
lang,
|
344 |
+
custom_tokenizer_file
|
345 |
)
|
346 |
|
347 |
if use_custom_embedding:
|
|
|
352 |
tokenizer, optimized_chunks = optimize_vocabulary(chunks)
|
353 |
chunks = optimized_chunks
|
354 |
|
|
|
355 |
results, search_time, vector_store = search_embeddings(
|
356 |
chunks,
|
357 |
embedding_model,
|
|
|
374 |
results_df = pd.DataFrame(all_results)
|
375 |
stats_df = pd.DataFrame(all_stats)
|
376 |
|
377 |
+
# Generate visualizations
|
378 |
+
fig = visualize_results(results_df, stats_df)
|
379 |
+
|
380 |
+
return results_df, stats_df, fig
|
381 |
|
382 |
def format_results(results, stats):
|
383 |
formatted_results = []
|
|
|
385 |
result = {
|
386 |
"Model": stats["model"],
|
387 |
"Content": doc.page_content,
|
388 |
+
"Embedding": doc.embedding if hasattr(doc, 'embedding') else None,
|
389 |
**doc.metadata,
|
390 |
**{k: v for k, v in stats.items() if k not in ["model"]}
|
391 |
}
|
392 |
formatted_results.append(result)
|
393 |
return formatted_results
|
394 |
|
395 |
+
# Gradio Interface
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
396 |
def launch_interface(share=True):
|
397 |
iface = gr.Interface(
|
398 |
fn=compare_embeddings,
|
399 |
inputs=[
|
400 |
gr.File(label="Upload File (Optional)"),
|
401 |
gr.Textbox(label="Search Query"),
|
402 |
+
gr.CheckboxGroup(choices=list(model_manager.list_models().keys()) + ["Custom"], label="Embedding Model Types"),
|
403 |
+
gr.CheckboxGroup(choices=[model for models in model_manager.list_models().values() for model in models] + ["custom_model"], label="Embedding Models"),
|
404 |
gr.Radio(choices=["token", "recursive"], label="Split Strategy", value="recursive"),
|
405 |
gr.Slider(100, 1000, step=100, value=500, label="Chunk Size"),
|
406 |
gr.Slider(0, 100, step=10, value=50, label="Overlap Size"),
|
|
|
411 |
gr.Dropdown(choices=["german", "english", "french"], label="Language", value="german"),
|
412 |
gr.Checkbox(label="Use Custom Embedding", value=False),
|
413 |
gr.Checkbox(label="Optimize Vocabulary", value=False),
|
414 |
+
gr.Slider(0, 1, step=0.1, value=0.3, label="Phonetic Matching Weight"),
|
415 |
+
gr.File(label="Custom Tokenizer File (Optional)")
|
416 |
],
|
417 |
outputs=[
|
418 |
gr.Dataframe(label="Results", interactive=False),
|
|
|
426 |
tutorial_md = """
|
427 |
# Advanced Embedding Comparison Tool Tutorial
|
428 |
|
429 |
+
This tool allows you to compare different embedding models and retrieval strategies for document search and similarity matching.
|
430 |
+
|
431 |
+
## How to use:
|
432 |
+
|
433 |
+
1. Upload a file (optional) or use the default files in the system.
|
434 |
+
2. Enter a search query.
|
435 |
+
3. Select one or more embedding model types and specific models.
|
436 |
+
4. Choose a text splitting strategy and set chunk size and overlap.
|
437 |
+
5. Select a vector store type and search type.
|
438 |
+
6. Set the number of top results to retrieve.
|
439 |
+
7. Choose the language of your documents.
|
440 |
+
8. Optionally, use custom embeddings, optimize vocabulary, or adjust phonetic matching weight.
|
441 |
+
9. If you have a custom tokenizer, upload the file.
|
442 |
+
|
443 |
+
The tool will process your query and display results, statistics, and visualizations to help you compare the performance of different models and strategies.
|
444 |
"""
|
445 |
|
446 |
iface = gr.TabbedInterface(
|