-
Notifications
You must be signed in to change notification settings - Fork 111
/
rag.py
555 lines (514 loc) · 22.3 KB
/
rag.py
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
# -*- coding: utf-8 -*-
"""
@author:XuMing([email protected])
@description:
"""
import argparse
import hashlib
import os
import re
from threading import Thread
from typing import Union, List
import jieba
import torch
from loguru import logger
from peft import PeftModel
from similarities import (
EnsembleSimilarity,
BertSimilarity,
BM25Similarity,
)
from similarities.similarity import SimilarityABC
from transformers import (
AutoModel,
AutoModelForCausalLM,
AutoTokenizer,
BloomForCausalLM,
BloomTokenizerFast,
LlamaTokenizer,
LlamaForCausalLM,
TextIteratorStreamer,
GenerationConfig,
AutoModelForSequenceClassification,
)
jieba.setLogLevel("ERROR")
MODEL_CLASSES = {
"bloom": (BloomForCausalLM, BloomTokenizerFast),
"chatglm": (AutoModel, AutoTokenizer),
"llama": (LlamaForCausalLM, LlamaTokenizer),
"baichuan": (AutoModelForCausalLM, AutoTokenizer),
"auto": (AutoModelForCausalLM, AutoTokenizer),
}
PROMPT_TEMPLATE = """基于以下已知信息,简洁和专业的来回答用户的问题。用简体中文回答。
已知内容:
{context_str}
问题:
{query_str}
"""
class SentenceSplitter:
def __init__(self, chunk_size: int = 250, chunk_overlap: int = 50):
self.chunk_size = chunk_size
self.chunk_overlap = chunk_overlap
def split_text(self, text: str) -> List[str]:
if self._is_has_chinese(text):
return self._split_chinese_text(text)
else:
return self._split_english_text(text)
def _split_chinese_text(self, text: str) -> List[str]:
sentence_endings = {'\n', '。', '!', '?', ';', '…'} # 句末标点符号
chunks, current_chunk = [], ''
for word in jieba.cut(text):
if len(current_chunk) + len(word) > self.chunk_size:
chunks.append(current_chunk.strip())
current_chunk = word
else:
current_chunk += word
if word[-1] in sentence_endings and len(current_chunk) > self.chunk_size - self.chunk_overlap:
chunks.append(current_chunk.strip())
current_chunk = ''
if current_chunk:
chunks.append(current_chunk.strip())
if self.chunk_overlap > 0 and len(chunks) > 1:
chunks = self._handle_overlap(chunks)
return chunks
def _split_english_text(self, text: str) -> List[str]:
# 使用正则表达式按句子分割英文文本
sentences = re.split(r'(?<=[.!?])\s+', text.replace('\n', ' '))
chunks = []
current_chunk = ''
for sentence in sentences:
if len(current_chunk) + len(sentence) <= self.chunk_size:
current_chunk += (' ' if current_chunk else '') + sentence
else:
if len(sentence) > self.chunk_size:
for i in range(0, len(sentence), self.chunk_size):
chunks.append(sentence[i:i + self.chunk_size])
current_chunk = ''
else:
chunks.append(current_chunk)
current_chunk = sentence
if current_chunk: # Add the last chunk
chunks.append(current_chunk)
if self.chunk_overlap > 0 and len(chunks) > 1:
chunks = self._handle_overlap(chunks)
return chunks
def _is_has_chinese(self, text: str) -> bool:
# check if contains chinese characters
if any("\u4e00" <= ch <= "\u9fff" for ch in text):
return True
else:
return False
def _handle_overlap(self, chunks: List[str]) -> List[str]:
# 处理块间重叠
overlapped_chunks = []
for i in range(len(chunks) - 1):
chunk = chunks[i] + ' ' + chunks[i + 1][:self.chunk_overlap]
overlapped_chunks.append(chunk.strip())
overlapped_chunks.append(chunks[-1])
return overlapped_chunks
class Rag:
def __init__(
self,
similarity_model: SimilarityABC = None,
generate_model_type: str = "auto",
generate_model_name_or_path: str = "Qwen/Qwen2-0.5B-Instruct",
lora_model_name_or_path: str = None,
corpus_files: Union[str, List[str]] = None,
save_corpus_emb_dir: str = "./corpus_embs/",
device: str = None,
int8: bool = False,
int4: bool = False,
chunk_size: int = 250,
chunk_overlap: int = 0,
rerank_model_name_or_path: str = None,
enable_history: bool = False,
num_expand_context_chunk: int = 2,
similarity_top_k: int = 10,
rerank_top_k: int = 3,
):
"""
Init RAG model.
:param similarity_model: similarity model, default None, if set, will use it instead of EnsembleSimilarity
:param generate_model_type: generate model type
:param generate_model_name_or_path: generate model name or path
:param lora_model_name_or_path: lora model name or path
:param corpus_files: corpus files
:param save_corpus_emb_dir: save corpus embeddings dir, default ./corpus_embs/
:param device: device, default None, auto select gpu or cpu
:param int8: use int8 quantization, default False
:param int4: use int4 quantization, default False
:param chunk_size: chunk size, default 250
:param chunk_overlap: chunk overlap, default 0, can not set to > 0 if num_expand_context_chunk > 0
:param rerank_model_name_or_path: rerank model name or path, default 'BAAI/bge-reranker-base'
:param enable_history: enable history, default False
:param num_expand_context_chunk: num expand context chunk, default 2, if set to 0, will not expand context chunk
:param similarity_top_k: similarity_top_k, default 5, similarity model search k corpus chunks
:param rerank_top_k: rerank_top_k, default 3, rerank model search k corpus chunks
"""
if torch.cuda.is_available():
default_device = torch.device(0)
elif torch.backends.mps.is_available():
default_device = torch.device('cpu')
else:
default_device = torch.device('cpu')
self.device = device or default_device
if num_expand_context_chunk > 0 and chunk_overlap > 0:
logger.warning(f" 'num_expand_context_chunk' and 'chunk_overlap' cannot both be greater than zero. "
f" 'chunk_overlap' has been set to zero by default.")
chunk_overlap = 0
self.text_splitter = SentenceSplitter(chunk_size, chunk_overlap)
if similarity_model is not None:
self.sim_model = similarity_model
else:
m1 = BertSimilarity(model_name_or_path="shibing624/text2vec-base-multilingual", device=self.device)
m2 = BM25Similarity()
default_sim_model = EnsembleSimilarity(similarities=[m1, m2], weights=[0.5, 0.5], c=2)
self.sim_model = default_sim_model
self.gen_model, self.tokenizer = self._init_gen_model(
generate_model_type,
generate_model_name_or_path,
peft_name=lora_model_name_or_path,
int8=int8,
int4=int4,
)
self.history = []
self.corpus_files = corpus_files
if corpus_files:
self.add_corpus(corpus_files)
self.save_corpus_emb_dir = save_corpus_emb_dir
if rerank_model_name_or_path is None:
rerank_model_name_or_path = "BAAI/bge-reranker-base"
if rerank_model_name_or_path:
self.rerank_tokenizer = AutoTokenizer.from_pretrained(rerank_model_name_or_path)
self.rerank_model = AutoModelForSequenceClassification.from_pretrained(rerank_model_name_or_path)
self.rerank_model.to(self.device)
self.rerank_model.eval()
else:
self.rerank_model = None
self.rerank_tokenizer = None
self.enable_history = enable_history
self.similarity_top_k = similarity_top_k
self.num_expand_context_chunk = num_expand_context_chunk
self.rerank_top_k = rerank_top_k
def __str__(self):
return f"Similarity model: {self.sim_model}, Generate model: {self.gen_model}"
def _init_gen_model(
self,
gen_model_type: str,
gen_model_name_or_path: str,
peft_name: str = None,
int8: bool = False,
int4: bool = False,
):
"""Init generate model."""
if int8 or int4:
device_map = None
else:
device_map = "auto"
model_class, tokenizer_class = MODEL_CLASSES[gen_model_type]
tokenizer = tokenizer_class.from_pretrained(gen_model_name_or_path, trust_remote_code=True)
model = model_class.from_pretrained(
gen_model_name_or_path,
load_in_8bit=int8 if gen_model_type not in ['baichuan', 'chatglm'] else False,
load_in_4bit=int4 if gen_model_type not in ['baichuan', 'chatglm'] else False,
torch_dtype="auto",
device_map=device_map,
trust_remote_code=True,
)
if self.device == torch.device('cpu'):
model.float()
if gen_model_type in ['baichuan', 'chatglm']:
if int4:
model = model.quantize(4).cuda()
elif int8:
model = model.quantize(8).cuda()
try:
model.generation_config = GenerationConfig.from_pretrained(gen_model_name_or_path, trust_remote_code=True)
except Exception as e:
logger.warning(f"Failed to load generation config from {gen_model_name_or_path}, {e}")
if peft_name:
model = PeftModel.from_pretrained(
model,
peft_name,
torch_dtype="auto",
)
logger.info(f"Loaded peft model from {peft_name}")
model.eval()
return model, tokenizer
def _get_chat_input(self):
messages = []
for conv in self.history:
if conv and len(conv) > 0 and conv[0]:
messages.append({'role': 'user', 'content': conv[0]})
if conv and len(conv) > 1 and conv[1]:
messages.append({'role': 'assistant', 'content': conv[1]})
input_ids = self.tokenizer.apply_chat_template(
conversation=messages,
tokenize=True,
add_generation_prompt=True,
return_tensors='pt'
)
return input_ids.to(self.gen_model.device)
@torch.inference_mode()
def stream_generate_answer(
self,
max_new_tokens=512,
temperature=0.7,
repetition_penalty=1.0,
context_len=2048
):
streamer = TextIteratorStreamer(self.tokenizer, timeout=60.0, skip_prompt=True, skip_special_tokens=True)
input_ids = self._get_chat_input()
max_src_len = context_len - max_new_tokens - 8
input_ids = input_ids[-max_src_len:]
generation_kwargs = dict(
input_ids=input_ids,
max_new_tokens=max_new_tokens,
temperature=temperature,
do_sample=True,
repetition_penalty=repetition_penalty,
streamer=streamer,
)
thread = Thread(target=self.gen_model.generate, kwargs=generation_kwargs)
thread.start()
yield from streamer
def add_corpus(self, files: Union[str, List[str]]):
"""Load document files."""
if isinstance(files, str):
files = [files]
for doc_file in files:
if doc_file.endswith('.pdf'):
corpus = self.extract_text_from_pdf(doc_file)
elif doc_file.endswith('.docx'):
corpus = self.extract_text_from_docx(doc_file)
elif doc_file.endswith('.md'):
corpus = self.extract_text_from_markdown(doc_file)
else:
corpus = self.extract_text_from_txt(doc_file)
full_text = '\n'.join(corpus)
chunks = self.text_splitter.split_text(full_text)
self.sim_model.add_corpus(chunks)
self.corpus_files = files
logger.debug(f"files: {files}, corpus size: {len(self.sim_model.corpus)}, top3: "
f"{list(self.sim_model.corpus.values())[:3]}")
@staticmethod
def get_file_hash(fpaths):
hasher = hashlib.md5()
target_file_data = bytes()
if isinstance(fpaths, str):
fpaths = [fpaths]
for fpath in fpaths:
with open(fpath, 'rb') as file:
chunk = file.read(1024 * 1024) # read only first 1MB
hasher.update(chunk)
target_file_data += chunk
hash_name = hasher.hexdigest()[:32]
return hash_name
@staticmethod
def extract_text_from_pdf(file_path: str):
"""Extract text content from a PDF file."""
import PyPDF2
contents = []
with open(file_path, 'rb') as f:
pdf_reader = PyPDF2.PdfReader(f)
for page in pdf_reader.pages:
page_text = page.extract_text().strip()
raw_text = [text.strip() for text in page_text.splitlines() if text.strip()]
new_text = ''
for text in raw_text:
new_text += text
if text[-1] in ['.', '!', '?', '。', '!', '?', '…', ';', ';', ':', ':', '”', '’', ')', '】', '》', '」',
'』', '〕', '〉', '》', '〗', '〞', '〟', '»', '"', "'", ')', ']', '}']:
contents.append(new_text)
new_text = ''
if new_text:
contents.append(new_text)
return contents
@staticmethod
def extract_text_from_txt(file_path: str):
"""Extract text content from a TXT file."""
with open(file_path, 'r', encoding='utf-8') as f:
contents = [text.strip() for text in f.readlines() if text.strip()]
return contents
@staticmethod
def extract_text_from_docx(file_path: str):
"""Extract text content from a DOCX file."""
import docx
document = docx.Document(file_path)
contents = [paragraph.text.strip() for paragraph in document.paragraphs if paragraph.text.strip()]
return contents
@staticmethod
def extract_text_from_markdown(file_path: str):
"""Extract text content from a Markdown file."""
import markdown
from bs4 import BeautifulSoup
with open(file_path, 'r', encoding='utf-8') as f:
markdown_text = f.read()
html = markdown.markdown(markdown_text)
soup = BeautifulSoup(html, 'html.parser')
contents = [text.strip() for text in soup.get_text().splitlines() if text.strip()]
return contents
@staticmethod
def _add_source_numbers(lst):
"""Add source numbers to a list of strings."""
return [f'[{idx + 1}]\t "{item}"' for idx, item in enumerate(lst)]
def _get_reranker_score(self, query: str, reference_results: List[str]):
"""Get reranker score."""
pairs = []
for reference in reference_results:
pairs.append([query, reference])
with torch.no_grad():
inputs = self.rerank_tokenizer(pairs, padding=True, truncation=True, return_tensors='pt', max_length=512)
inputs_on_device = {k: v.to(self.rerank_model.device) for k, v in inputs.items()}
scores = self.rerank_model(**inputs_on_device, return_dict=True).logits.view(-1, ).float()
return scores
def get_reference_results(self, query: str):
"""
Get reference results.
1. Similarity model get similar chunks
2. Rerank similar chunks
3. Expand reference context chunk
:param query:
:return:
"""
reference_results = []
sim_contents = self.sim_model.most_similar(query, topn=self.similarity_top_k)
# Get reference results from corpus
hit_chunk_dict = dict()
for c in sim_contents:
for id_score_dict in c:
corpus_id = id_score_dict['corpus_id']
hit_chunk = id_score_dict["corpus_doc"]
reference_results.append(hit_chunk)
hit_chunk_dict[corpus_id] = hit_chunk
if reference_results:
if self.rerank_model is not None:
# Rerank reference results
rerank_scores = self._get_reranker_score(query, reference_results)
logger.debug(f"rerank_scores: {rerank_scores}")
# Get rerank top k chunks
reference_results = [reference for reference, score in sorted(
zip(reference_results, rerank_scores), key=lambda x: x[1], reverse=True)][:self.rerank_top_k]
hit_chunk_dict = {corpus_id: hit_chunk for corpus_id, hit_chunk in hit_chunk_dict.items() if
hit_chunk in reference_results}
# Expand reference context chunk
if self.num_expand_context_chunk > 0:
new_reference_results = []
for corpus_id, hit_chunk in hit_chunk_dict.items():
expanded_reference = self.sim_model.corpus.get(corpus_id - 1, '') + hit_chunk
for i in range(self.num_expand_context_chunk):
expanded_reference += self.sim_model.corpus.get(corpus_id + i + 1, '')
new_reference_results.append(expanded_reference)
reference_results = new_reference_results
return reference_results
def predict_stream(
self,
query: str,
max_length: int = 512,
context_len: int = 2048,
temperature: float = 0.7,
):
"""Generate predictions stream."""
stop_str = self.tokenizer.eos_token if self.tokenizer.eos_token else "</s>"
if not self.enable_history:
self.history = []
if self.sim_model.corpus:
reference_results = self.get_reference_results(query)
if reference_results:
reference_results = self._add_source_numbers(reference_results)
context_str = '\n'.join(reference_results)[:(context_len - len(PROMPT_TEMPLATE))]
else:
context_str = ''
prompt = PROMPT_TEMPLATE.format(context_str=context_str, query_str=query)
else:
prompt = query
logger.debug(f"prompt: {prompt}")
self.history.append([prompt, ''])
response = ""
for new_text in self.stream_generate_answer(
max_new_tokens=max_length,
temperature=temperature,
context_len=context_len,
):
if new_text != stop_str:
response += new_text
yield response
def predict(
self,
query: str,
max_length: int = 512,
context_len: int = 2048,
temperature: float = 0.7,
):
"""Query from corpus."""
reference_results = []
if not self.enable_history:
self.history = []
if self.sim_model.corpus:
reference_results = self.get_reference_results(query)
if reference_results:
reference_results = self._add_source_numbers(reference_results)
context_str = '\n'.join(reference_results)[:(context_len - len(PROMPT_TEMPLATE))]
else:
context_str = ''
prompt = PROMPT_TEMPLATE.format(context_str=context_str, query_str=query)
else:
prompt = query
logger.debug(f"prompt: {prompt}")
self.history.append([prompt, ''])
response = ""
for new_text in self.stream_generate_answer(
max_new_tokens=max_length,
temperature=temperature,
context_len=context_len,
):
response += new_text
response = response.strip()
self.history[-1][1] = response
return response, reference_results
def query(self, query: str, **kwargs):
return self.predict(query, **kwargs)
def save_corpus_emb(self):
dir_name = self.get_file_hash(self.corpus_files)
save_dir = os.path.join(self.save_corpus_emb_dir, dir_name)
if hasattr(self.sim_model, 'save_corpus_embeddings'):
self.sim_model.save_corpus_embeddings(save_dir)
logger.debug(f"Saving corpus embeddings to {save_dir}")
return save_dir
def load_corpus_emb(self, emb_dir: str):
if hasattr(self.sim_model, 'load_corpus_embeddings'):
logger.debug(f"Loading corpus embeddings from {emb_dir}")
self.sim_model.load_corpus_embeddings(emb_dir)
if __name__ == "__main__":
parser = argparse.ArgumentParser()
parser.add_argument("--sim_model_name", type=str, default="shibing624/text2vec-base-multilingual")
parser.add_argument("--gen_model_type", type=str, default="auto")
parser.add_argument("--gen_model_name", type=str, default="Qwen/Qwen2-0.5B-Instruct")
parser.add_argument("--lora_model", type=str, default=None)
parser.add_argument("--rerank_model_name", type=str, default="")
parser.add_argument("--corpus_files", type=str, default="data/sample.pdf")
parser.add_argument("--device", type=str, default=None)
parser.add_argument("--int4", action='store_true', help="use int4 quantization")
parser.add_argument("--int8", action='store_true', help="use int8 quantization")
parser.add_argument("--chunk_size", type=int, default=220)
parser.add_argument("--chunk_overlap", type=int, default=0)
parser.add_argument("--num_expand_context_chunk", type=int, default=1)
args = parser.parse_args()
print(args)
sim_model = BertSimilarity(model_name_or_path=args.sim_model_name, device=args.device)
m = Rag(
similarity_model=sim_model,
generate_model_type=args.gen_model_type,
generate_model_name_or_path=args.gen_model_name,
lora_model_name_or_path=args.lora_model,
device=args.device,
int4=args.int4,
int8=args.int8,
chunk_size=args.chunk_size,
chunk_overlap=args.chunk_overlap,
corpus_files=args.corpus_files.split(','),
num_expand_context_chunk=args.num_expand_context_chunk,
rerank_model_name_or_path=args.rerank_model_name,
)
r, refs = m.predict('自然语言中的非平行迁移是指什么?')
print(r, refs)