File size: 57,070 Bytes
ccb8edf e9d5e9c 8ad24b7 2618588 8ad24b7 84cb849 8ad24b7 a5f7e3b 8ad24b7 950a593 23e19ad a967d35 950a593 cc69ccc c118719 cc69ccc cd94d30 5b5f687 5480240 5b5f687 efa7515 5b5f687 689088d ceb7f25 efa7515 5b5f687 efa7515 8ad24b7 488e992 54a0f5c a8ae89c 488e992 8ad24b7 b35adb8 8ad24b7 488e992 8ad24b7 ea0ce95 8ad24b7 ea0ce95 8ad24b7 a5caef8 a14da67 ea1c5b5 a14da67 a5caef8 ea0ce95 a5caef8 ea0ce95 c62b613 a967d35 c62b613 ea0ce95 ce988dc 8ad24b7 c77f8ac 8ad24b7 c77f8ac 8ad24b7 c77f8ac 8ad24b7 950a593 8ad24b7 60de941 b35adb8 60de941 8ad24b7 60de941 8ad24b7 950a593 8ad24b7 b35adb8 8ad24b7 b35adb8 8ad24b7 60de941 b35adb8 8ad24b7 c77f8ac 8ad24b7 60de941 8ad24b7 b35adb8 c38e61c 60de941 8ad24b7 b35adb8 8ad24b7 60de941 b35adb8 8ad24b7 187e575 52d5b06 950a593 52d5b06 950a593 52d5b06 8ad24b7 c38e61c 8ad24b7 52d5b06 c09985f 8ad24b7 60de941 8ad24b7 950a593 c38e61c 8ad24b7 a8ae89c e6d71ec c38e61c b35adb8 8ad24b7 b35adb8 8ad24b7 c38e61c 8ad24b7 54a0f5c 8ad24b7 54a0f5c 8ad24b7 738ada4 8ad24b7 a717449 0646ad5 027365f 0646ad5 a5f7e3b 0646ad5 d78ad1e 950a593 8ad24b7 c38e61c 84cb849 9efbb97 8ad24b7 f9d798a 8ad24b7 f9d798a 950a593 9efbb97 84cb849 f9d798a c77f8ac 9efbb97 84cb849 8ad24b7 ea0ce95 c77f8ac 84cb849 8ad24b7 950a593 d78ad1e 950a593 8ad24b7 84cb849 8ad24b7 c38e61c 8ad24b7 ea0ce95 8ad24b7 84cb849 ea0ce95 950a593 31095a2 950a593 1078648 c38e61c 84cb849 54a0f5c 9efbb97 84cb849 e0a87f8 1078648 950a593 c38e61c 1078648 9efbb97 84cb849 9efbb97 1754322 8ad24b7 2577444 fed8ef0 9efbb97 fed8ef0 9efbb97 a8ae89c 8ad24b7 9efbb97 fed8ef0 6fd2acf cc69ccc 6fd2acf cc69ccc 6fd2acf cc69ccc 6fd2acf a14da67 6fd2acf cc69ccc 6fd2acf cc69ccc 6fd2acf cc69ccc 6fd2acf aef3987 6fd2acf aef3987 6fd2acf aef3987 cc69ccc 6fd2acf aef3987 cc69ccc 6fd2acf cc69ccc aef3987 6fd2acf caacbc1 c1aaeb5 caacbc1 c1aaeb5 cc69ccc a14da67 a5caef8 ea1c5b5 a14da67 a5caef8 cc69ccc 825e760 cc69ccc c1aaeb5 cc69ccc c1aaeb5 cc69ccc 8ad24b7 950a593 c38e61c ea0ce95 950a593 cc69ccc 950a593 ea0ce95 950a593 ea0ce95 950a593 ea0ce95 950a593 f0f9414 6fd2acf cc69ccc 6fd2acf cc69ccc 586cc45 cc69ccc 586cc45 eafad45 586cc45 c38e61c 586cc45 cc69ccc 586cc45 6fd2acf cc69ccc 6fd2acf ea0ce95 f0f9414 8ad24b7 f0f9414 8ad24b7 f0f9414 8ad24b7 f0f9414 8ad24b7 f0f9414 8ad24b7 f0f9414 a967d35 f0f9414 6fa8e54 f0f9414 c38e61c f0f9414 8ad24b7 e7e5bc4 f0f9414 8ad24b7 4e5a67d 6fd2acf a5caef8 6fd2acf a5caef8 6fd2acf 950a593 |
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360 361 362 363 364 365 366 367 368 369 370 371 372 373 374 375 376 377 378 379 380 381 382 383 384 385 386 387 388 389 390 391 392 393 394 395 396 397 398 399 400 401 402 403 404 405 406 407 408 409 410 411 412 413 414 415 416 417 418 419 420 421 422 423 424 425 426 427 428 429 430 431 432 433 434 435 436 437 438 439 440 441 442 443 444 445 446 447 448 449 450 451 452 453 454 455 456 457 458 459 460 461 462 463 464 465 466 467 468 469 470 471 472 473 474 475 476 477 478 479 480 481 482 483 484 485 486 487 488 489 490 491 492 493 494 495 496 497 498 499 500 501 502 503 504 505 506 507 508 509 510 511 512 513 514 515 516 517 518 519 520 521 522 523 524 525 526 527 528 529 530 531 532 533 534 535 536 537 538 539 540 541 542 543 544 545 546 547 548 549 550 551 552 553 554 555 556 557 558 559 560 561 562 563 564 565 566 567 568 569 570 571 572 573 574 575 576 577 578 579 580 581 582 583 584 585 586 587 588 589 590 591 592 593 594 595 596 597 598 599 600 601 602 603 604 605 606 607 608 609 610 611 612 613 614 615 616 617 618 619 620 621 622 623 624 625 626 627 628 629 630 631 632 633 634 635 636 637 638 639 640 641 642 643 644 645 646 647 648 649 650 651 652 653 654 655 656 657 658 659 660 661 662 663 664 665 666 667 668 669 670 671 672 673 674 675 676 677 678 679 680 681 682 683 684 685 686 687 688 689 690 691 692 693 694 695 696 697 698 699 700 701 702 703 704 705 706 707 708 709 710 711 712 713 714 715 716 717 718 719 720 721 722 723 724 725 726 727 728 729 730 731 732 733 734 735 736 737 738 739 740 741 742 743 744 745 746 747 748 749 750 751 752 753 754 755 756 757 758 759 760 761 762 763 764 765 766 767 768 769 770 771 772 773 774 775 776 777 778 779 780 781 782 783 784 785 786 787 788 789 790 791 792 793 794 795 796 797 798 799 800 801 802 803 804 805 806 807 808 809 810 811 812 813 814 815 816 817 818 819 820 821 822 823 824 825 826 827 828 829 830 831 832 833 834 835 836 837 838 839 840 841 842 843 844 845 846 847 848 849 850 851 852 853 854 855 856 857 858 859 860 861 862 863 864 865 866 867 868 869 870 871 872 873 874 875 876 877 878 879 880 881 882 883 884 885 886 887 888 889 890 891 892 893 894 895 896 897 898 899 900 901 902 903 904 905 906 907 908 909 910 911 912 913 914 915 916 917 918 919 920 921 922 923 924 925 926 927 928 929 930 931 932 933 934 935 936 937 938 939 940 941 942 943 944 945 946 947 948 949 950 951 952 953 954 955 956 957 958 959 960 961 962 963 964 965 966 967 968 969 970 971 972 973 974 975 976 977 978 979 980 981 982 983 984 985 986 987 988 989 990 991 992 993 994 995 996 997 998 999 1000 1001 1002 1003 1004 1005 1006 1007 1008 1009 1010 1011 1012 1013 1014 1015 1016 1017 1018 1019 1020 1021 1022 1023 1024 1025 1026 1027 1028 1029 1030 1031 1032 1033 1034 1035 1036 1037 1038 1039 1040 1041 1042 1043 1044 1045 1046 1047 1048 1049 1050 1051 1052 1053 1054 1055 1056 1057 1058 1059 1060 1061 1062 1063 1064 1065 1066 1067 1068 1069 1070 1071 1072 1073 1074 1075 1076 1077 1078 1079 1080 1081 1082 1083 1084 1085 1086 1087 1088 1089 1090 1091 1092 1093 1094 1095 1096 1097 1098 1099 1100 1101 1102 1103 1104 1105 1106 1107 1108 1109 1110 1111 1112 1113 1114 1115 1116 1117 1118 1119 1120 1121 1122 1123 1124 1125 1126 1127 1128 1129 1130 1131 1132 1133 1134 1135 1136 1137 1138 1139 1140 1141 1142 1143 1144 1145 1146 1147 1148 1149 1150 1151 1152 1153 1154 1155 1156 1157 1158 1159 1160 1161 1162 1163 1164 1165 1166 1167 1168 1169 1170 1171 1172 1173 1174 1175 1176 1177 1178 1179 1180 1181 1182 1183 1184 1185 1186 1187 1188 1189 1190 1191 1192 1193 1194 1195 1196 1197 1198 1199 1200 1201 1202 1203 1204 1205 1206 1207 1208 1209 1210 1211 1212 1213 1214 1215 1216 1217 1218 1219 1220 1221 1222 1223 1224 1225 1226 1227 1228 1229 1230 1231 1232 1233 1234 1235 1236 1237 1238 1239 1240 1241 1242 1243 1244 1245 1246 1247 1248 1249 1250 1251 1252 1253 1254 1255 1256 1257 1258 1259 1260 1261 1262 1263 1264 1265 1266 1267 1268 1269 1270 1271 1272 1273 1274 1275 1276 1277 1278 1279 1280 1281 1282 1283 1284 1285 1286 1287 1288 1289 1290 1291 1292 1293 1294 1295 1296 1297 1298 1299 1300 1301 1302 1303 1304 1305 1306 1307 1308 1309 1310 1311 1312 1313 1314 1315 1316 1317 1318 1319 1320 1321 1322 1323 1324 1325 1326 1327 1328 1329 1330 1331 1332 1333 1334 1335 1336 1337 1338 1339 1340 1341 1342 1343 1344 1345 1346 1347 1348 1349 1350 1351 1352 1353 1354 1355 1356 1357 |
import os
import time
import pdfplumber
import docx
import nltk
import gradio as gr
from langchain_huggingface import HuggingFaceEmbeddings
from langchain_community.embeddings import CohereEmbeddings
from langchain_openai import OpenAIEmbeddings
from langchain_community.vectorstores import FAISS, Chroma
from langchain_text_splitters import RecursiveCharacterTextSplitter, TokenTextSplitter
from typing import List, Dict, Any
import pandas as pd
import numpy as np
import re
from nltk.corpus import stopwords
from nltk.tokenize import word_tokenize
from nltk.stem import SnowballStemmer
import jellyfish
from gensim.models import Word2Vec
from gensim.models.fasttext import FastText
from collections import Counter
from tokenizers import Tokenizer, models, trainers
from tokenizers.models import WordLevel
from tokenizers.trainers import WordLevelTrainer
from tokenizers.pre_tokenizers import Whitespace
import matplotlib.pyplot as plt
import seaborn as sns
from sklearn.manifold import TSNE
from sklearn.metrics import silhouette_score
from scipy.stats import spearmanr
from functools import lru_cache
from langchain.retrievers import MultiQueryRetriever
from langchain_huggingface.llms import HuggingFacePipeline
from langchain.schema import NoOpRunManager
from transformers import pipeline
from sklearn.model_selection import ParameterGrid
from tqdm import tqdm
import random
from huggingface_hub import login
hf_token = os.getenv("hf_token")
login(token=hf_token)
# Define the model pipeline with additional generation parameters
model_pipeline = pipeline(
model="meta-llama/Llama-3.2-1B",
#use_auth_token=hf_token,
max_length=1000, # You can increase this if needed
max_new_tokens=500 # Limit how many tokens are generated
)
# Use the pipeline in HuggingFacePipeline
llm = HuggingFacePipeline(pipeline=model_pipeline)
# NLTK Resource Download
def download_nltk_resources():
resources = ['punkt', 'stopwords', 'snowball_data']
for resource in resources:
try:
nltk.download(resource, quiet=True)
except Exception as e:
print(f"Failed to download {resource}: {str(e)}")
download_nltk_resources()
nltk.download('punkt')
FILES_DIR = './files'
# Model Management
class ModelManager:
def __init__(self):
self.models = {
'HuggingFace': {
'e5-base-de': "danielheinz/e5-base-sts-en-de",
'paraphrase-miniLM': "paraphrase-multilingual-MiniLM-L12-v2",
'paraphrase-mpnet': "paraphrase-multilingual-mpnet-base-v2",
'gte-large': "gte-large",
'gbert-base': "gbert-base"
},
'OpenAI': {
'text-embedding-ada-002': "text-embedding-ada-002"
},
'Cohere': {
'embed-multilingual-v2.0': "embed-multilingual-v2.0"
}
}
def add_model(self, provider, name, model_path):
if provider not in self.models:
self.models[provider] = {}
self.models[provider][name] = model_path
def remove_model(self, provider, name):
if provider in self.models and name in self.models[provider]:
del self.models[provider][name]
def get_model(self, provider, name):
return self.models.get(provider, {}).get(name)
def list_models(self):
return {provider: list(models.keys()) for provider, models in self.models.items()}
model_manager = ModelManager()
# File Handling
class FileHandler:
@staticmethod
def extract_text(file_path):
ext = os.path.splitext(file_path)[-1].lower()
if ext == '.pdf':
return FileHandler._extract_from_pdf(file_path)
elif ext == '.docx':
return FileHandler._extract_from_docx(file_path)
elif ext == '.txt':
return FileHandler._extract_from_txt(file_path)
else:
raise ValueError(f"Unsupported file type: {ext}")
@staticmethod
def _extract_from_pdf(file_path):
with pdfplumber.open(file_path) as pdf:
return ' '.join([page.extract_text() for page in pdf.pages])
@staticmethod
def _extract_from_docx(file_path):
doc = docx.Document(file_path)
return ' '.join([para.text for para in doc.paragraphs])
@staticmethod
def _extract_from_txt(file_path):
with open(file_path, 'r', encoding='utf-8') as f:
return f.read()
# Text Processing
def simple_tokenize(text):
return text.split()
def preprocess_text(text, lang='german', apply_preprocessing=True):
if not apply_preprocessing:
return text
text = text.lower()
text = re.sub(r'[^a-zA-Z\s]', '', text)
try:
tokens = word_tokenize(text, language=lang)
except LookupError:
print(f"Warning: NLTK punkt tokenizer for {lang} not found. Using simple tokenization.")
tokens = simple_tokenize(text)
try:
stop_words = set(stopwords.words(lang))
except LookupError:
print(f"Warning: Stopwords for {lang} not found. Skipping stopword removal.")
stop_words = set()
tokens = [token for token in tokens if token not in stop_words]
try:
stemmer = SnowballStemmer(lang)
tokens = [stemmer.stem(token) for token in tokens]
except ValueError:
print(f"Warning: SnowballStemmer for {lang} not available. Skipping stemming.")
return ' '.join(tokens)
def phonetic_match(text, query, method='levenshtein_distance', apply_phonetic=True):
if not apply_phonetic:
return 0
if method == 'levenshtein_distance':
text_phonetic = jellyfish.soundex(text)
query_phonetic = jellyfish.soundex(query)
return jellyfish.levenshtein_distance(text_phonetic, query_phonetic)
return 0
#def optimize_query(query, llm_model):
def optimize_query(query, llm_model, chunks, embedding_model, vector_store_type, search_type, top_k):
# Use a HuggingFace model for text generation
#model_id = "google/flan-t5-large"
#tokenizer = AutoTokenizer.from_pretrained(model_id)
#model = AutoModelForCausalLM.from_pretrained(model_id)
#pipe = pipeline(
# "text-generation", model=model, tokenizer=tokenizer, max_new_tokens=512
#)
#llm = HuggingFacePipeline(pipeline=pipe)
#llm = HuggingFacePipeline(pipeline(model="HuggingFaceH4/zephyr-7b-beta"))
# Create a temporary vector store for query optimization
temp_vector_store = get_vector_store(vector_store_type, chunks, embedding_model)
# Create a retriever with the temporary vector store
temp_retriever = get_retriever(temp_vector_store, search_type, {"k": top_k})
multi_query_retriever = MultiQueryRetriever.from_llm(
retriever=temp_retriever,
llm=llm
)
# Use a NoOpRunManager as the run_manager
optimized_queries = multi_query_retriever.invoke(query)
return optimized_queries
def create_custom_embedding(texts, model_type='word2vec', vector_size=100, window=5, min_count=1):
tokenized_texts = [text.split() for text in texts]
if model_type == 'word2vec':
model = Word2Vec(sentences=tokenized_texts, vector_size=vector_size, window=window, min_count=min_count, workers=4)
elif model_type == 'fasttext':
model = FastText(sentences=tokenized_texts, vector_size=vector_size, window=window, min_count=min_count, workers=4)
else:
raise ValueError("Unsupported model type")
return model
class CustomEmbeddings(HuggingFaceEmbeddings):
def __init__(self, model_path):
self.model = Word2Vec.load(model_path) # or FastText.load() for FastText models
def embed_documents(self, texts):
return [self.model.wv[text.split()] for text in texts]
def embed_query(self, text):
return self.model.wv[text.split()]
# Custom Tokenizer
def create_custom_tokenizer(file_path, model_type='WordLevel', vocab_size=10000, special_tokens=None):
with open(file_path, 'r', encoding='utf-8') as f:
text = f.read()
if model_type == 'WordLevel':
tokenizer = Tokenizer(WordLevel(unk_token="[UNK]"))
elif model_type == 'BPE':
tokenizer = Tokenizer(models.BPE(unk_token="[UNK]"))
elif model_type == 'Unigram':
tokenizer = Tokenizer(models.Unigram())
else:
raise ValueError(f"Unsupported tokenizer model: {model_type}")
tokenizer.pre_tokenizer = Whitespace()
special_tokens = special_tokens or ["[UNK]", "[CLS]", "[SEP]", "[PAD]", "[MASK]"]
trainer = trainers.WordLevelTrainer(special_tokens=special_tokens, vocab_size=vocab_size)
tokenizer.train_from_iterator([text], trainer)
return tokenizer
def custom_tokenize(text, tokenizer):
return tokenizer.encode(text).tokens
# Embedding and Vector Store
#@lru_cache(maxsize=None)
# Helper functions
def get_text_splitter(split_strategy, chunk_size, overlap_size, custom_separators=None):
if split_strategy == 'token':
return TokenTextSplitter(chunk_size=chunk_size, chunk_overlap=overlap_size)
elif split_strategy == 'recursive':
return RecursiveCharacterTextSplitter(
chunk_size=chunk_size,
chunk_overlap=overlap_size,
separators=custom_separators or ["\n\n", "\n", " ", ""]
)
else:
raise ValueError(f"Unsupported split strategy: {split_strategy}")
def get_embedding_model(model_type, model_name):
model_path = model_manager.get_model(model_type, model_name)
if model_type == 'HuggingFace':
return HuggingFaceEmbeddings(model_name=model_path)
elif model_type == 'OpenAI':
return OpenAIEmbeddings(model=model_path)
elif model_type == 'Cohere':
return CohereEmbeddings(model=model_path)
else:
raise ValueError(f"Unsupported model type: {model_type}")
def get_vector_store(vector_store_type, chunks, embedding_model):
chunks_tuple = tuple(chunks)
if vector_store_type == 'FAISS':
return FAISS.from_texts(chunks, embedding_model)
elif vector_store_type == 'Chroma':
return Chroma.from_texts(chunks, embedding_model)
else:
raise ValueError(f"Unsupported vector store type: {vector_store_type}")
def get_retriever(vector_store, search_type, search_kwargs):
if search_type == 'similarity':
return vector_store.as_retriever(search_type="similarity", search_kwargs=search_kwargs)
elif search_type == 'mmr':
return vector_store.as_retriever(search_type="mmr", search_kwargs=search_kwargs)
elif search_type == 'custom':
return vector_store.as_retriever(search_type="similarity", search_kwargs=search_kwargs)
else:
raise ValueError(f"Unsupported search type: {search_type}")
def custom_similarity(query_embedding, doc_embedding, query, doc_text, phonetic_weight=0.3):
embedding_sim = np.dot(query_embedding, doc_embedding) / (np.linalg.norm(query_embedding) * np.linalg.norm(doc_embedding))
phonetic_sim = phonetic_match(doc_text, query)
combined_sim = (1 - phonetic_weight) * embedding_sim + phonetic_weight * phonetic_sim
return combined_sim
def _create_vector_store(vector_store_type, chunks_tuple, embedding_model):
chunks = list(chunks_tuple)
if vector_store_type == 'FAISS':
return FAISS.from_texts(chunks, embedding_model)
elif vector_store_type == 'Chroma':
return Chroma.from_texts(chunks, embedding_model)
else:
raise ValueError(f"Unsupported vector store type: {vector_store_type}")
# Main Processing Functions
def process_files(file_path, model_type, model_name, split_strategy, chunk_size, overlap_size, custom_separators, lang='german', apply_preprocessing=True, custom_tokenizer_file=None, custom_tokenizer_model=None, custom_tokenizer_vocab_size=10000, custom_tokenizer_special_tokens=None):
if file_path:
text = FileHandler.extract_text(file_path)
else:
text = ""
for file in os.listdir(FILES_DIR):
file_path = os.path.join(FILES_DIR, file)
text += FileHandler.extract_text(file_path)
if custom_tokenizer_file:
tokenizer = create_custom_tokenizer(custom_tokenizer_file, custom_tokenizer_model, custom_tokenizer_vocab_size, custom_tokenizer_special_tokens)
text = ' '.join(custom_tokenize(text, tokenizer))
elif apply_preprocessing:
text = preprocess_text(text, lang)
text_splitter = get_text_splitter(split_strategy, chunk_size, overlap_size, custom_separators)
chunks = text_splitter.split_text(text)
embedding_model = get_embedding_model(model_type, model_name)
return chunks, embedding_model, len(text.split())
def search_embeddings(chunks, embedding_model, vector_store_type, search_type, query, top_k, expected_result=None, lang='german', apply_phonetic=True, phonetic_weight=0.3):
preprocessed_query = preprocess_text(query, lang) if apply_phonetic else query
vector_store = get_vector_store(vector_store_type, chunks, embedding_model)
retriever = get_retriever(vector_store, search_type, {"k": top_k})
start_time = time.time()
results = retriever.invoke(preprocessed_query)
def score_result(doc):
similarity_score = vector_store.similarity_search_with_score(doc.page_content, k=1)[0][1]
if apply_phonetic:
phonetic_score = phonetic_match(doc.page_content, query)
return (1 - phonetic_weight) * similarity_score + phonetic_weight * phonetic_score
else:
return similarity_score
results = sorted(results, key=score_result, reverse=True)
end_time = time.time()
embeddings = []
for doc in results:
if hasattr(doc, 'embedding'):
embeddings.append(doc.embedding)
else:
embeddings.append(None)
results_df = pd.DataFrame({
'content': [doc.page_content for doc in results],
'embedding': embeddings,
'length': [len(doc.page_content) for doc in results],
'contains_expected': [expected_result in doc.page_content if expected_result else None for doc in results]
})
return results_df, end_time - start_time, vector_store, results
# Evaluation Metrics
# ... (previous code remains the same)
def calculate_statistics(results, search_time, vector_store, num_tokens, embedding_model, query, top_k, expected_result=None):
stats = {
"num_results": len(results),
"avg_content_length": np.mean([len(doc.page_content) for doc in results]) if results else 0,
"min_content_length": min([len(doc.page_content) for doc in results]) if results else 0,
"max_content_length": max([len(doc.page_content) for doc in results]) if results else 0,
"search_time": search_time,
"vector_store_size": vector_store._index.ntotal if hasattr(vector_store, '_index') else "N/A",
"num_documents": len(vector_store.docstore._dict),
"num_tokens": num_tokens,
"embedding_vocab_size": embedding_model.client.get_vocab_size() if hasattr(embedding_model, 'client') and hasattr(embedding_model.client, 'get_vocab_size') else "N/A",
"embedding_dimension": len(embedding_model.embed_query(query)),
"top_k": top_k,
}
if expected_result:
stats["contains_expected"] = any(expected_result in doc.page_content for doc in results)
stats["expected_result_rank"] = next((i for i, doc in enumerate(results) if expected_result in doc.page_content), -1) + 1
if len(results) > 1000:
embeddings = [embedding_model.embed_query(doc.page_content) for doc in results]
pairwise_similarities = np.inner(embeddings, embeddings)
stats["result_diversity"] = 1 - np.mean(pairwise_similarities[np.triu_indices(len(embeddings), k=1)])
if len(embeddings) > 2:
stats["silhouette_score"] = silhouette_score(embeddings, range(len(embeddings)))
else:
stats["silhouette_score"] = "N/A"
else:
stats["result_diversity"] = "N/A"
stats["silhouette_score"] = "N/A"
query_embedding = embedding_model.embed_query(query)
result_embeddings = [embedding_model.embed_query(doc.page_content) for doc in results]
similarities = [np.inner(query_embedding, emb) for emb in result_embeddings]
rank_correlation, _ = spearmanr(similarities, range(len(similarities)))
stats["rank_correlation"] = rank_correlation
return stats
# Visualization
def visualize_results(results_df, stats_df):
fig, axs = plt.subplots(2, 2, figsize=(20, 20))
sns.barplot(x='model', y='search_time', data=stats_df, ax=axs[0, 0])
axs[0, 0].set_title('Search Time by Model')
axs[0, 0].set_xticks(range(len(axs[0, 0].get_xticklabels())))
axs[0, 0].set_xticklabels(axs[0, 0].get_xticklabels(), rotation=45, ha='right')
sns.scatterplot(x='result_diversity', y='rank_correlation', hue='model', data=stats_df, ax=axs[0, 1])
axs[0, 1].set_title('Result Diversity vs. Rank Correlation')
sns.boxplot(x='model', y='avg_content_length', data=stats_df, ax=axs[1, 0])
axs[1, 0].set_title('Distribution of Result Content Lengths')
axs[1, 0].set_xticks(range(len(axs[0, 0].get_xticklabels())))
axs[1, 0].set_xticklabels(axs[1, 0].get_xticklabels(), rotation=45, ha='right')
embeddings = np.array([embedding for embedding in results_df['embedding'] if isinstance(embedding, np.ndarray)])
if len(embeddings) > 1:
tsne = TSNE(n_components=2, random_state=42)
embeddings_2d = tsne.fit_transform(embeddings)
sns.scatterplot(x=embeddings_2d[:, 0], y=embeddings_2d[:, 1], hue=results_df['model'][:len(embeddings)], ax=axs[1, 1])
axs[1, 1].set_title('t-SNE Visualization of Result Embeddings')
else:
axs[1, 1].text(0.5, 0.5, "Not enough data for t-SNE visualization", ha='center', va='center')
plt.tight_layout()
return fig
def optimize_vocabulary(texts, vocab_size=10000, min_frequency=2):
tokenizer = Tokenizer(models.BPE(unk_token="[UNK]"))
word_freq = Counter(word for text in texts for word in text.split())
optimized_texts = [
' '.join(word for word in text.split() if word_freq[word] >= min_frequency)
for text in texts
]
trainer = trainers.BpeTrainer(vocab_size=vocab_size, special_tokens=["[UNK]", "[CLS]", "[SEP]", "[PAD]", "[MASK]"])
tokenizer.train_from_iterator(optimized_texts, trainer)
return tokenizer, optimized_texts
# New postprocessing function
def rerank_results(results, query, reranker):
reranked_results = reranker.rerank(query, [doc.page_content for doc in results])
return reranked_results
# Main Comparison Function
def compare_embeddings(file, query, embedding_models, custom_embedding_model, split_strategy, chunk_size, overlap_size, custom_separators, vector_store_type, search_type, top_k, expected_result=None, lang='german', apply_preprocessing=True, optimize_vocab=False, apply_phonetic=True, phonetic_weight=0.3, custom_tokenizer_file=None, custom_tokenizer_model=None, custom_tokenizer_vocab_size=10000, custom_tokenizer_special_tokens=None, use_query_optimization=False, query_optimization_model="google/flan-t5-base", use_reranking=False):
all_results = []
all_stats = []
settings = {
"split_strategy": split_strategy,
"chunk_size": chunk_size,
"overlap_size": overlap_size,
"custom_separators": custom_separators,
"vector_store_type": vector_store_type,
"search_type": search_type,
"top_k": top_k,
"lang": lang,
"apply_preprocessing": apply_preprocessing,
"optimize_vocab": optimize_vocab,
"apply_phonetic": apply_phonetic,
"phonetic_weight": phonetic_weight,
"use_query_optimization": use_query_optimization,
"use_reranking": use_reranking
}
# Parse the embedding models from the checkbox group
models = [model.split(':') for model in embedding_models]
if custom_embedding_model:
models.append(custom_embedding_model.strip().split(':'))
for model_type, model_name in models:
chunks, embedding_model, num_tokens = process_files(
file.name if file else None,
model_type,
model_name,
split_strategy,
chunk_size,
overlap_size,
custom_separators.split(',') if custom_separators else None,
lang,
apply_preprocessing,
custom_tokenizer_file,
custom_tokenizer_model,
int(custom_tokenizer_vocab_size),
custom_tokenizer_special_tokens.split(',') if custom_tokenizer_special_tokens else None
)
if optimize_vocab:
tokenizer, optimized_chunks = optimize_vocabulary(chunks)
chunks = optimized_chunks
if use_query_optimization:
optimized_queries = optimize_query(query, query_optimization_model, chunks, embedding_model, vector_store_type, search_type, top_k)
query = " ".join(optimized_queries)
results, search_time, vector_store, results_raw = search_embeddings(
chunks,
embedding_model,
vector_store_type,
search_type,
query,
top_k,
expected_result,
lang,
apply_phonetic,
phonetic_weight
)
if use_reranking:
reranker = pipeline("text-classification", model="cross-encoder/ms-marco-MiniLM-L-12-v2")
results_raw = rerank_results(results_raw, query, reranker)
result_embeddings = [doc.metadata.get('embedding', None) for doc in results_raw]
stats = calculate_statistics(results_raw, search_time, vector_store, num_tokens, embedding_model, query, top_k, expected_result)
stats["model"] = f"{model_type} - {model_name}"
stats["model_type"] = model_type
stats["model_name"] = model_name
stats.update(settings)
formatted_results = format_results(results_raw, stats)
for i, result in enumerate(formatted_results):
result['embedding'] = result_embeddings[i]
result['length'] = len(result['Content'])
result['contains_expected'] = expected_result in result['Content'] if expected_result else None
all_results.extend(formatted_results)
all_stats.append(stats)
results_df = pd.DataFrame(all_results)
stats_df = pd.DataFrame(all_stats)
fig = visualize_results(results_df, stats_df)
best_results = analyze_results(stats_df)
return results_df, stats_df, fig, best_results
def format_results(results, stats):
formatted_results = []
for doc in results:
result = {
"Model": stats["model"],
"Content": doc.page_content,
"Embedding": doc.embedding if hasattr(doc, 'embedding') else None,
**doc.metadata,
**{k: v for k, v in stats.items() if k not in ["model"]}
}
formatted_results.append(result)
return formatted_results
#####
from sklearn.model_selection import ParameterGrid
from tqdm import tqdm
# ... (previous code remains the same)
# function for automated testing
def automated_testing(file, query, test_params, expected_result=None):
all_results = []
all_stats = []
param_grid = ParameterGrid(test_params)
for params in tqdm(param_grid, desc="Running tests"):
chunks, embedding_model, num_tokens = process_files(
file.name if file else None,
params['model_type'],
params['model_name'],
params['split_strategy'],
params['chunk_size'],
params['overlap_size'],
params.get('custom_separators', None),
params['lang'],
params['apply_preprocessing'],
params.get('custom_tokenizer_file', None),
params.get('custom_tokenizer_model', None),
params.get('custom_tokenizer_vocab_size', 10000),
params.get('custom_tokenizer_special_tokens', None)
)
if params['optimize_vocab']:
tokenizer, optimized_chunks = optimize_vocabulary(chunks)
chunks = optimized_chunks
if params['use_query_optimization']:
optimized_queries = optimize_query(query, params['query_optimization_model'])
query = " ".join(optimized_queries)
results, search_time, vector_store, results_raw = search_embeddings(
chunks,
embedding_model,
params['vector_store_type'],
params['search_type'],
query,
params['top_k'],
expected_result,
params['lang'],
params['apply_phonetic'],
params['phonetic_weight']
)
if params['use_reranking']:
reranker = pipeline("text-classification", model="cross-encoder/ms-marco-MiniLM-L-12-v2")
results_raw = rerank_results(results_raw, query, reranker)
stats = calculate_statistics(results_raw, search_time, vector_store, num_tokens, embedding_model, query, params['top_k'], expected_result)
stats["model"] = f"{params['model_type']} - {params['model_name']}"
stats["model_type"] = params['model_type']
stats["model_name"] = params['model_name']
stats.update(params)
all_results.extend(format_results(results_raw, stats))
all_stats.append(stats)
return pd.DataFrame(all_results), pd.DataFrame(all_stats)
# Function to analyze results and propose best model and settings
def analyze_results(stats_df):
metric_weights = {
'search_time': -0.3,
'result_diversity': 0.2,
'rank_correlation': 0.3,
'silhouette_score': 0.2,
'contains_expected': 0.5, # High weight for containing the expected result
'expected_result_rank': -0.4 # Lower rank (closer to 1) is better
}
for metric in metric_weights.keys():
stats_df[metric] = pd.to_numeric(stats_df[metric], errors='coerce')
stats_df['weighted_score'] = sum(
stats_df[metric].fillna(0) * weight
for metric, weight in metric_weights.items()
)
best_config = stats_df.loc[stats_df['weighted_score'].idxmax()]
recommendations = {
'best_model': f"{best_config['model_type']} - {best_config['model_name']}",
'best_settings': {
'split_strategy': best_config['split_strategy'],
'chunk_size': int(best_config['chunk_size']),
'overlap_size': int(best_config['overlap_size']),
'vector_store_type': best_config['vector_store_type'],
'search_type': best_config['search_type'],
'top_k': int(best_config['top_k']),
'optimize_vocab': bool(best_config['optimize_vocab']),
'use_query_optimization': bool(best_config['use_query_optimization']),
'use_reranking': bool(best_config['use_reranking']),
'lang': best_config['lang'],
'apply_preprocessing': bool(best_config['apply_preprocessing']),
'apply_phonetic': bool(best_config['apply_phonetic']),
'phonetic_weight': float(best_config['phonetic_weight'])
},
'performance_summary': {
'search_time': float(best_config['search_time']),
'result_diversity': float(best_config['result_diversity']),
'rank_correlation': float(best_config['rank_correlation']),
'silhouette_score': float(best_config['silhouette_score']),
'contains_expected': bool(best_config['contains_expected']),
'expected_result_rank': int(best_config['expected_result_rank'])
}
}
return recommendations
####
def get_llm_suggested_settings(file, num_chunks=1):
if not file:
return {"error": "No file uploaded"}
chunks, _, _ = process_files(
file.name,
'HuggingFace',
'paraphrase-miniLM',
'recursive',
250,
50,
custom_separators=None
)
# Select a few random chunks
sample_chunks = random.sample(chunks, min(num_chunks, len(chunks)))
# Prepare the prompt
prompt = f"""Given the following text chunks from a document, suggest optimal settings for an embedding-based search system. The settings should include:
1. Embedding model type and name
2. Split strategy (token or recursive)
3. Chunk size
4. Overlap size
5. Vector store type (FAISS or Chroma)
6. Search type (similarity, mmr, or custom)
7. Top K results to retrieve
8. Whether to apply preprocessing
9. Whether to optimize vocabulary
10. Whether to apply phonetic matching
Text chunks:
{' '.join(sample_chunks)}
Provide your suggestions in a Python dictionary format."""
# Use a HuggingFace model for text generation
#model_id = "google/flan-t5-large"
#tokenizer = AutoTokenizer.from_pretrained(model_id)
#model = AutoModelForCausalLM.from_pretrained(model_id)
#pipe = pipeline(
# "text-generation", model=model, tokenizer=tokenizer, max_new_tokens=512
#)
#llm = HuggingFacePipeline(pipeline=pipe)
#llm = HuggingFacePipeline(pipeline(model="HuggingFaceH4/zephyr-7b-beta"))
#llm = HuggingFacePipeline.from_model_id(
# model_id="google/flan-t5-large",
# task="text2text-generation",
# model_kwargs={"do_sample": True, "temperature": 0.7, "max_new_tokens": 512},
#)
# Generate suggestions
suggested_settings = llm.invoke(prompt)
# Parse the generated text to extract the dictionary
try:
settings_dict = eval(suggested_settings)
# Convert the settings to match the interface inputs
return {
"embedding_models": f"{settings_dict['embedding_model_type']}:{settings_dict['embedding_model_name']}",
"split_strategy": settings_dict["split_strategy"],
"chunk_size": settings_dict["chunk_size"],
"overlap_size": settings_dict["overlap_size"],
"vector_store_type": settings_dict["vector_store_type"],
"search_type": settings_dict["search_type"],
"top_k": settings_dict["top_k"],
"apply_preprocessing": settings_dict["apply_preprocessing"],
"optimize_vocab": settings_dict["optimize_vocabulary"],
"apply_phonetic": settings_dict["apply_phonetic_matching"],
"phonetic_weight": 0.3 # Default value, as it's not in the LLM suggestions
}
except:
return {"error": "Failed to parse LLM suggestions"}
def update_inputs_with_llm_suggestions(suggestions):
if suggestions is None or "error" in suggestions:
return [gr.update() for _ in range(11)] # Return no updates if there's an error or None
return [
gr.update(value=[suggestions["embedding_models"]]), # embedding_models_input
gr.update(value=suggestions["split_strategy"]), # split_strategy_input
gr.update(value=suggestions["chunk_size"]), # chunk_size_input
gr.update(value=suggestions["overlap_size"]), # overlap_size_input
gr.update(value=suggestions["vector_store_type"]), # vector_store_type_input
gr.update(value=suggestions["search_type"]), # search_type_input
gr.update(value=suggestions["top_k"]), # top_k_input
gr.update(value=suggestions["apply_preprocessing"]), # apply_preprocessing_input
gr.update(value=suggestions["optimize_vocab"]), # optimize_vocab_input
gr.update(value=suggestions["apply_phonetic"]), # apply_phonetic_input
gr.update(value=suggestions["phonetic_weight"]) # phonetic_weight_input
]
# Gradio Interface
def launch_interface(share=True):
with gr.Blocks() as iface:
gr.Markdown("# Advanced Embedding Comparison Tool")
with gr.Tab("Simple"):
file_input = gr.File(label="Upload File (Optional)")
query_input = gr.Textbox(label="Search Query")
expected_result_input = gr.Textbox(label="Expected Result (Optional)")
embedding_models_input = gr.CheckboxGroup(
choices=[
"HuggingFace:paraphrase-miniLM",
"HuggingFace:paraphrase-mpnet",
"OpenAI:text-embedding-ada-002",
"Cohere:embed-multilingual-v2.0"
],
label="Embedding Models"
)
top_k_input = gr.Slider(1, 10, step=1, value=5, label="Top K")
with gr.Tab("Advanced"):
custom_embedding_model_input = gr.Textbox(label="Custom Embedding Model (optional, format: type:name)")
split_strategy_input = gr.Radio(choices=["token", "recursive"], label="Split Strategy", value="recursive")
chunk_size_input = gr.Slider(100, 1000, step=100, value=500, label="Chunk Size")
overlap_size_input = gr.Slider(0, 100, step=10, value=50, label="Overlap Size")
custom_separators_input = gr.Textbox(label="Custom Split Separators (comma-separated, optional)")
vector_store_type_input = gr.Radio(choices=["FAISS", "Chroma"], label="Vector Store Type", value="FAISS")
search_type_input = gr.Radio(choices=["similarity", "mmr", "custom"], label="Search Type", value="similarity")
lang_input = gr.Dropdown(choices=["german", "english", "french"], label="Language", value="german")
with gr.Tab("Optional"):
apply_preprocessing_input = gr.Checkbox(label="Apply Text Preprocessing", value=True)
optimize_vocab_input = gr.Checkbox(label="Optimize Vocabulary", value=False)
apply_phonetic_input = gr.Checkbox(label="Apply Phonetic Matching", value=True)
phonetic_weight_input = gr.Slider(0, 1, step=0.1, value=0.3, label="Phonetic Matching Weight")
custom_tokenizer_file_input = gr.File(label="Custom Tokenizer File (Optional)")
custom_tokenizer_model_input = gr.Textbox(label="Custom Tokenizer Model (e.g., WordLevel, BPE, Unigram)")
custom_tokenizer_vocab_size_input = gr.Textbox(label="Custom Tokenizer Vocab Size", value="10000")
custom_tokenizer_special_tokens_input = gr.Textbox(label="Custom Tokenizer Special Tokens (comma-separated)")
use_query_optimization_input = gr.Checkbox(label="Use Query Optimization", value=False)
query_optimization_model_input = gr.Textbox(label="Query Optimization Model", value="google/flan-t5-base")
use_reranking_input = gr.Checkbox(label="Use Reranking", value=False)
with gr.Tab("Automation"):
auto_file_input = gr.File(label="Upload File (Optional)")
auto_query_input = gr.Textbox(label="Search Query")
auto_expected_result_input = gr.Textbox(label="Expected Result (Optional)")
auto_model_types = gr.CheckboxGroup(
choices=["HuggingFace", "OpenAI", "Cohere"],
label="Model Types to Test"
)
auto_model_names = gr.TextArea(label="Model Names to Test (comma-separated)")
auto_split_strategies = gr.CheckboxGroup(
choices=["token", "recursive"],
label="Split Strategies to Test"
)
auto_chunk_sizes = gr.TextArea(label="Chunk Sizes to Test (comma-separated)")
auto_overlap_sizes = gr.TextArea(label="Overlap Sizes to Test (comma-separated)")
auto_vector_store_types = gr.CheckboxGroup(
choices=["FAISS", "Chroma"],
label="Vector Store Types to Test"
)
auto_search_types = gr.CheckboxGroup(
choices=["similarity", "mmr", "custom"],
label="Search Types to Test"
)
auto_top_k = gr.TextArea(label="Top K Values to Test (comma-separated)")
auto_optimize_vocab = gr.Checkbox(label="Test Vocabulary Optimization", value=True)
auto_use_query_optimization = gr.Checkbox(label="Test Query Optimization", value=True)
auto_use_reranking = gr.Checkbox(label="Test Reranking", value=True)
with gr.Tab("LLM Suggestions"):
llm_file_input = gr.File(label="Upload File for LLM Suggestions")
llm_num_chunks = gr.Slider(1, 10, step=1, value=5, label="Number of Sample Chunks")
llm_suggest_button = gr.Button("Get LLM Suggestions")
llm_suggestions_output = gr.JSON(label="LLM-suggested Settings")
llm_suggest_button.click(
fn=get_llm_suggested_settings,
inputs=[llm_file_input, llm_num_chunks],
outputs=[llm_suggestions_output]
).then(
fn=update_inputs_with_llm_suggestions,
inputs=[llm_suggestions_output],
outputs=[
embedding_models_input, split_strategy_input, chunk_size_input,
overlap_size_input, vector_store_type_input, search_type_input,
top_k_input, apply_preprocessing_input, optimize_vocab_input,
apply_phonetic_input, phonetic_weight_input
]
)
results_output = gr.Dataframe(label="Results", interactive=False)
stats_output = gr.Dataframe(label="Statistics", interactive=False)
plot_output = gr.Plot(label="Visualizations")
best_settings_output = gr.JSON(label="Best Settings")
submit_button = gr.Button("Compare Embeddings")
submit_button.click(
#fn=lambda *args: compare_and_show_best(*args),
fn=lambda *args: compare_embeddings(*args),
inputs=[
file_input, query_input, embedding_models_input, custom_embedding_model_input,
split_strategy_input, chunk_size_input, overlap_size_input, custom_separators_input,
vector_store_type_input, search_type_input, top_k_input, expected_result_input, lang_input,
apply_preprocessing_input, optimize_vocab_input, apply_phonetic_input,
phonetic_weight_input, custom_tokenizer_file_input, custom_tokenizer_model_input,
custom_tokenizer_vocab_size_input, custom_tokenizer_special_tokens_input,
use_query_optimization_input, query_optimization_model_input, use_reranking_input
],
outputs=[results_output, stats_output, plot_output, best_settings_output]
)
auto_results_output = gr.Dataframe(label="Automated Test Results", interactive=False)
auto_stats_output = gr.Dataframe(label="Automated Test Statistics", interactive=False)
recommendations_output = gr.JSON(label="Recommendations")
auto_submit_button = gr.Button("Run Automated Tests")
auto_submit_button.click(
fn=lambda *args: run_automated_tests_and_analyze(*args),
inputs=[
auto_file_input, auto_query_input, auto_expected_result_input, auto_model_types, auto_model_names,
auto_split_strategies, auto_chunk_sizes, auto_overlap_sizes,
auto_vector_store_types, auto_search_types, auto_top_k,
auto_optimize_vocab, auto_use_query_optimization, auto_use_reranking
],
outputs=[auto_results_output, auto_stats_output, recommendations_output]
)
###
use_case_md = """
# 🚀 AI Act Embedding Use Case Guide
## 📚 Use Case: Embedding the German AI Act for Local Chat Retrieval
In this guide, we'll walk through the process of embedding the German version of the AI Act using our advanced embedding tool and MTEB. We'll then use these embeddings in a local chat application as a retriever/context.
### Step 1: Prepare the Document 📄
1. Download the German version of the AI Act (let's call it `ai_act_de.txt`).
2. Place the file in your project directory.
### Step 2: Set Up the Embedding Tool 🛠️
1. Open the Embedding Comparison Tool.
2. Navigate to the new "Automation" tab.
### Step 3: Configure the Automated Test 🔧
In the "Use Case" tab, set up the following configuration:
```markdown
- File: ai_act_de.txt
- Query: "Wie definiert das Gesetz KI-Systeme?"
- Model Types: ✅ HuggingFace, ✅ Sentence Transformers
- Model Names: paraphrase-multilingual-MiniLM-L12-v2, distiluse-base-multilingual-cased-v2
- Split Strategies: ✅ recursive, ✅ token
- Chunk Sizes: 256, 512, 1024
- Overlap Sizes: 32, 64, 128
- Vector Store Types: ✅ FAISS
- Search Types: ✅ similarity, ✅ mmr
- Top K Values: 3, 5, 7
- Test Vocabulary Optimization: ✅
- Test Query Optimization: ✅
- Test Reranking: ✅
```
### Step 4: Run the Automated Test 🏃♂️
Click the "Run Automated Tests" button and wait for the results.
### Step 5: Analyze the Results 📊
Let's say we got the following simulated results:
```markdown
Best Model: Sentence Transformers - paraphrase-multilingual-MiniLM-L12-v2
Best Settings:
- Split Strategy: recursive
- Chunk Size: 512
- Overlap Size: 64
- Vector Store Type: FAISS
- Search Type: mmr
- Top K: 5
- Optimize Vocabulary: True
- Use Query Optimization: True
- Use Reranking: True
Performance Summary:
- Search Time: 0.15s
- Result Diversity: 0.82
- Rank Correlation: 0.91
- Silhouette Score: 0.76
```
### Step 6: Understand the Results 🧠
1. **Model**: The Sentence Transformers model performed better, likely due to its multilingual capabilities and fine-tuning for paraphrasing tasks.
2. **Split Strategy**: Recursive splitting worked best, probably because it respects the document's structure better than fixed-length token splitting.
3. **Chunk Size**: 512 tokens provide a good balance between context and specificity.
4. **Search Type**: MMR (Maximum Marginal Relevance) outperformed simple similarity search, likely due to its ability to balance relevance and diversity in results.
5. **Optimizations**: All optimizations (vocabulary, query, and reranking) proved beneficial, indicating that the extra processing time is worth the improved results.
### Step 7: Implement in Local Chat 💬
Now that we have the optimal settings, let's implement this in a local chat application:
1. Use the `paraphrase-multilingual-MiniLM-L12-v2` model for embeddings.
2. Set up a FAISS vector store with the embedded chunks.
3. Implement MMR search with a top-k of 5.
4. Include the optimization steps in your pipeline.
### Step 8: Test the Implementation 🧪
Create a simple chat interface and test with various queries about the AI Act. For example:
User: "Was sind die Hauptziele des KI-Gesetzes?"
"""
tutorial_md = """
# Advanced Embedding Comparison Tool Tutorial
Welcome to the **Advanced Embedding Comparison Tool**! This comprehensive guide will help you understand and utilize the tool's features to optimize your **Retrieval-Augmented Generation (RAG)** systems.
## Table of Contents
1. [Introduction to RAG](#introduction-to-rag)
2. [Key Components of RAG](#key-components-of-rag)
3. [Impact of Parameter Changes](#impact-of-parameter-changes)
4. [Advanced Features](#advanced-features)
5. [Using the Embedding Comparison Tool](#using-the-embedding-comparison-tool)
6. [Automated Testing and Analysis](#automated-testing-and-analysis)
7. [Mathematical Concepts and Metrics](#mathematical-concepts-and-metrics)
8. [Code Examples](#code-examples)
9. [Best Practices and Tips](#best-practices-and-tips)
10. [Resources and Further Reading](#resources-and-further-reading)
---
## Introduction to RAG
**Retrieval-Augmented Generation (RAG)** is a powerful technique that combines the strengths of large language models (LLMs) with the ability to access and use external knowledge. RAG is particularly useful for:
- Providing up-to-date information
- Answering questions based on specific documents or data sources
- Reducing hallucinations in AI responses
- Customizing AI outputs for specific domains or use cases
RAG is ideal for applications requiring accurate, context-specific information retrieval combined with natural language generation, such as chatbots, question-answering systems, and document analysis tools.
---
## Key Components of RAG
### 1. Document Loading
Ingests documents from various sources (PDFs, web pages, databases, etc.) into a format that can be processed by the RAG system. The tool supports multiple file formats, including PDF, DOCX, and TXT.
### 2. Document Splitting
Splits large documents into smaller chunks for more efficient processing and retrieval. Available strategies include:
- **Token-based splitting**
- **Recursive splitting**
### 3. Vector Store and Embeddings
Embeddings are dense vector representations of text that capture semantic meaning. The tool supports multiple embedding models and vector stores:
- **Embedding models**: HuggingFace, OpenAI, Cohere, and custom models.
- **Vector stores**: FAISS and Chroma.
### 4. Retrieval
Finds the most relevant documents or chunks based on a query. Available retrieval methods include:
- **Similarity search**
- **Maximum Marginal Relevance (MMR)**
- **Custom search methods**
---
## Impact of Parameter Changes
Understanding how different parameters affect your RAG system is crucial for optimization:
- **Chunk Size**: Larger chunks provide more context but may reduce precision. Smaller chunks increase precision but may lose context.
- **Overlap**: More overlap helps maintain context between chunks but increases computational load.
- **Embedding Model**: Performance varies across languages and domains.
- **Vector Store**: Affects query speed and the types of searches.
- **Retrieval Method**: Influences the diversity and relevance of retrieved documents.
---
## Advanced Features
### 1. Custom Tokenization
Upload a custom tokenizer file and specify the tokenizer model, vocabulary size, and special tokens for domain or language-specific tokenization.
### 2. Query Optimization
Improve search results by generating multiple variations of the input query using a language model to capture different phrasings.
### 3. Reranking
Further refine search results by using a separate model to re-score and reorder the initial retrieval results.
### 4. Phonetic Matching
For languages like German, phonetic matching with adjustable weighting is available.
### 5. Vocabulary Optimization
Optimize vocabulary for domain-specific applications during the embedding process.
---
## Using the Embedding Comparison Tool
The tool is divided into several tabs for ease of use:
### Simple Tab
1. **File Upload**: Upload a file (PDF, DOCX, or TXT) or use files from the `./files` directory.
2. **Search Query**: Enter the search query.
3. **Embedding Models**: Select one or more embedding models to compare.
4. **Top K**: Set the number of top results to retrieve (1-10).
### Advanced Tab
5. **Custom Embedding Model**: Specify a custom embedding model.
6. **Split Strategy**: Choose between 'token' and 'recursive' splitting.
7. **Chunk Size**: Set chunk size (100-1000).
8. **Overlap Size**: Set overlap between chunks (0-100).
9. **Custom Split Separators**: Enter custom separators for text splitting.
10. **Vector Store Type**: Choose between FAISS and Chroma.
11. **Search Type**: Select 'similarity', 'mmr', or 'custom'.
12. **Language**: Specify the document's primary language.
### Optional Tab
13. **Text Preprocessing**: Toggle text preprocessing.
14. **Vocabulary Optimization**: Enable vocabulary optimization.
15. **Phonetic Matching**: Enable phonetic matching and set its weight.
16. **Custom Tokenizer**: Upload a custom tokenizer and specify parameters.
17. **Query Optimization**: Enable query optimization and specify the model.
18. **Reranking**: Enable result reranking.
---
## Automated Testing and Analysis
The **Automation tab** allows you to run comprehensive tests across multiple configurations:
1. Set up test parameters like model types, split strategies, chunk sizes, etc.
2. Click "Run Automated Tests."
3. View results, statistics, and recommendations to find optimal configurations for your use case.
---
## Mathematical Concepts and Metrics
### Cosine Similarity
Measures the cosine of the angle between two vectors, used in similarity search:
$$\text{cosine similarity} = \frac{\mathbf{A} \cdot \mathbf{B}}{\|\mathbf{A}\| \|\mathbf{B}\|}$$
### Maximum Marginal Relevance (MMR)
Balances relevance and diversity in search results:
$$\text{MMR} = \arg\max_{D_i \in R \setminus S} [\lambda \text{Sim}_1(D_i, Q) - (1-\lambda) \max_{D_j \in S} \text{Sim}_2(D_i, D_j)]$$
### Silhouette Score
Measures how well an object fits within its own cluster compared to others. Scores range from -1 to 1, where higher values indicate better-defined clusters.
---
## Code Examples
### Custom Tokenization
```python
def create_custom_tokenizer(file_path, model_type='WordLevel', vocab_size=10000, special_tokens=None):
with open(file_path, 'r', encoding='utf-8') as f:
text = f.read()
tokenizer = Tokenizer(models.WordLevel(unk_token="[UNK]")) if model_type == 'WordLevel' else Tokenizer(models.BPE(unk_token="[UNK]"))
tokenizer.pre_tokenizer = Whitespace()
trainer = trainers.WordLevelTrainer(special_tokens=special_tokens or ["[UNK]", "[CLS]", "[SEP]", "[PAD]", "[MASK]"], vocab_size=vocab_size)
tokenizer.train_from_iterator([text], trainer)
return tokenizer
````
### Query Optimization
```python
def optimize_query(query, llm):
multi_query_retriever = MultiQueryRetriever.from_llm(
retriever=get_retriever(vector_store, search_type, search_kwargs),
llm=llm
)
optimized_queries = multi_query_retriever.invoke(query)
return optimized_queries
````
### Reranking
```python
def rerank_results(results, query, reranker):
reranked_results = reranker.rerank(query, [doc.page_content for doc in results])
return reranked_results
````
### Best Practices and Tips
- Start Simple: Begin with basic configurations, then gradually add complexity.
- Benchmark: Use automated testing to benchmark different setups.
- Domain-Specific Tuning: Consider custom tokenizers and embeddings for specialized domains.
- Balance Performance and Cost: Use advanced features like query optimization and reranking judiciously.
- Iterate: Optimization is an iterative process—refine your approach based on tool insights.
## Useful Resources and Links
Here are some valuable resources to help you better understand and work with embeddings, retrieval systems, and natural language processing:
### Embeddings and Vector Databases
- [Understanding Embeddings](https://www.tensorflow.org/text/guide/word_embeddings): A guide by TensorFlow on word embeddings
- [FAISS: A Library for Efficient Similarity Search](https://github.com/facebookresearch/faiss): Facebook AI's vector similarity search library
- [Chroma: The AI-native open-source embedding database](https://www.trychroma.com/): An embedding database designed for AI applications
### Natural Language Processing
- [NLTK (Natural Language Toolkit)](https://www.nltk.org/): A leading platform for building Python programs to work with human language data
- [spaCy](https://spacy.io/): Industrial-strength Natural Language Processing in Python
- [Hugging Face Transformers](https://huggingface.co/transformers/): State-of-the-art Natural Language Processing for PyTorch and TensorFlow 2.0
### Retrieval-Augmented Generation (RAG)
- [LangChain](https://python.langchain.com/docs/get_started/introduction): A framework for developing applications powered by language models
- [OpenAI's RAG Tutorial](https://platform.openai.com/docs/tutorials/web-qa-embeddings): A guide on building a QA system with embeddings
### German Language Processing
- [Kölner Phonetik](https://en.wikipedia.org/wiki/Cologne_phonetics): Information about the Kölner Phonetik algorithm
- [German NLP Resources](https://github.com/adbar/German-NLP): A curated list of open-access resources for German NLP
### Benchmarks and Evaluation
- [MTEB Leaderboard](https://huggingface.co/spaces/mteb/leaderboard): Massive Text Embedding Benchmark leaderboard
- [GLUE Benchmark](https://gluebenchmark.com/): General Language Understanding Evaluation benchmark
### Tools and Libraries
- [Gensim](https://radimrehurek.com/gensim/): Topic modelling for humans
- [Sentence-Transformers](https://www.sbert.net/): A Python framework for state-of-the-art sentence, text and image embeddings
### Support me
- [Visual Crew Builder](https://visual-crew.builder.ai/): Tool for create AI systems, workflows and api. Or just a notebook.
This tool empowers you to fine-tune your RAG system for optimal performance. Experiment with different settings, run automated tests, and use insights to create an efficient information retrieval and generation system.
# Template
python
´´´
# Chat App Template
def create_chat_app(settings):
def chat(message, history):
# Process the message using the configured embedding model and vector store
chunks, embedding_model, _ = process_files(
settings['file_path'],
settings['model_type'],
settings['model_name'],
settings['split_strategy'],
settings['chunk_size'],
settings['overlap_size'],
settings['custom_separators'],
settings['lang'],
settings['apply_preprocessing']
)
results, _, _, _ = search_embeddings(
chunks,
embedding_model,
settings['vector_store_type'],
settings['search_type'],
message,
settings['top_k'],
lang=settings['lang'],
apply_phonetic=settings['apply_phonetic'],
phonetic_weight=settings['phonetic_weight']
)
# Generate a response based on the retrieved results
response = f"Based on the query '{message}', here are the top {settings['top_k']} relevant results:\n\n"
for i, result in enumerate(results[:settings['top_k']]):
response += f"{i+1}. {result['content'][:100]}...\n\n"
return response
with gr.Blocks() as chat_interface:
gr.Markdown(f"# Chat App using {settings['model_type']} - {settings['model_name']}")
chatbot = gr.Chatbot()
msg = gr.Textbox()
clear = gr.Button("Clear")
msg.submit(chat, [msg, chatbot], [msg, chatbot])
clear.click(lambda: None, None, chatbot, queue=False)
return chat_interface
# Sample usage of the chat app template
sample_settings = {
'file_path': 'path/to/your/document.pdf',
'model_type': 'HuggingFace',
'model_name': 'paraphrase-miniLM',
'split_strategy': 'recursive',
'chunk_size': 500,
'overlap_size': 50,
'custom_separators': None,
'vector_store_type': 'FAISS',
'search_type': 'similarity',
'top_k': 3,
'lang': 'english',
'apply_preprocessing': True,
'apply_phonetic': True,
'phonetic_weight': 0.3
}
sample_chat_app = create_chat_app(sample_settings)
if __name__ == "__main__":
launch_interface()
# Uncomment the following line to launch the sample chat app
´´´
"""
iface = gr.TabbedInterface(
[iface, gr.Markdown(tutorial_md), gr.Markdown( use_case_md )],
["Embedding Comparison", "Tutorial", "Use Case"]
)
iface.launch(share=share)
def run_automated_tests_and_analyze(*args):
file, query, auto_expected_result_input, model_types, model_names, split_strategies, chunk_sizes, overlap_sizes, \
vector_store_types, search_types, top_k_values, optimize_vocab, use_query_optimization, use_reranking = args
test_params = {
'model_type': model_types,
'model_name': [name.strip() for name in model_names.split(',')],
'split_strategy': split_strategies,
'chunk_size': [int(size.strip()) for size in chunk_sizes.split(',')],
'overlap_size': [int(size.strip()) for size in overlap_sizes.split(',')],
'vector_store_type': vector_store_types,
'search_type': search_types,
'top_k': [int(k.strip()) for k in top_k_values.split(',')],
'lang': ['german'], # You can expand this if needed
'apply_preprocessing': [True],
'optimize_vocab': [optimize_vocab],
'apply_phonetic': [True],
'phonetic_weight': [0.3],
'use_query_optimization': [use_query_optimization],
'query_optimization_model': ['google/flan-t5-base'],
'use_reranking': [use_reranking]
}
results_df, stats_df = automated_testing(file, query, test_params, auto_expected_result_input)
recommendations = analyze_results(stats_df)
return results_df, stats_df, recommendations
if __name__ == "__main__":
launch_interface() |