Spaces:
Sleeping
Sleeping
ZoniaChatbot
commited on
Commit
•
8730d86
1
Parent(s):
05e3b86
Upload chatpdf.py
Browse files- chatpdf.py +582 -0
chatpdf.py
ADDED
@@ -0,0 +1,582 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
import argparse
|
2 |
+
import hashlib
|
3 |
+
import os
|
4 |
+
import re
|
5 |
+
from threading import Thread
|
6 |
+
from typing import Union, List
|
7 |
+
|
8 |
+
import jieba
|
9 |
+
import torch
|
10 |
+
from loguru import logger
|
11 |
+
from peft import PeftModel
|
12 |
+
from similarities import (
|
13 |
+
EnsembleSimilarity,
|
14 |
+
BertSimilarity,
|
15 |
+
BM25Similarity,
|
16 |
+
)
|
17 |
+
from similarities.similarity import SimilarityABC
|
18 |
+
from transformers import (
|
19 |
+
AutoModelForCausalLM,
|
20 |
+
AutoTokenizer,
|
21 |
+
TextIteratorStreamer,
|
22 |
+
GenerationConfig,
|
23 |
+
AutoModelForSequenceClassification,
|
24 |
+
)
|
25 |
+
|
26 |
+
jieba.setLogLevel("ERROR")
|
27 |
+
|
28 |
+
MODEL_CLASSES = {
|
29 |
+
"auto": (AutoModelForCausalLM, AutoTokenizer),
|
30 |
+
}
|
31 |
+
|
32 |
+
PROMPT_TEMPLATE1 = """Utiliza la siguiente información para responder a la pregunta del usuario.
|
33 |
+
Si no sabes la respuesta, di simplemente que no la sabes, no intentes inventarte una respuesta.
|
34 |
+
|
35 |
+
Contexto: {context_str}
|
36 |
+
Pregunta: {query_str}
|
37 |
+
|
38 |
+
Devuelve sólo la respuesta útil que aparece a continuación y nada más, y ésta debe estar en Español.
|
39 |
+
Respuesta útil:
|
40 |
+
"""
|
41 |
+
PROMPT_TEMPLATE = """Basándose en la siguiente información conocida, responda a la pregunta del usuario de forma
|
42 |
+
concisa y profesional. Si no puede obtener una respuesta, diga «No se puede responder a la pregunta basándose en la
|
43 |
+
información conocida» o «No se proporciona suficiente información relevante», no está permitido añadir elementos
|
44 |
+
inventados en la respuesta.
|
45 |
+
|
46 |
+
Contenido conocido:
|
47 |
+
{context_str}
|
48 |
+
|
49 |
+
Pregunta:
|
50 |
+
{query_str}
|
51 |
+
"""
|
52 |
+
|
53 |
+
|
54 |
+
class SentenceSplitter:
|
55 |
+
def __init__(self, chunk_size: int = 250, chunk_overlap: int = 50):
|
56 |
+
self.chunk_size = chunk_size
|
57 |
+
self.chunk_overlap = chunk_overlap
|
58 |
+
|
59 |
+
def split_text(self, text: str) -> List[str]:
|
60 |
+
if self._is_has_chinese(text):
|
61 |
+
return self._split_chinese_text(text)
|
62 |
+
else:
|
63 |
+
return self._split_english_text(text)
|
64 |
+
|
65 |
+
def _split_chinese_text(self, text: str) -> List[str]:
|
66 |
+
sentence_endings = {'\n', '。', '!', '?', ';', '…'} # puntuación al final de una frase
|
67 |
+
chunks, current_chunk = [], ''
|
68 |
+
for word in jieba.cut(text):
|
69 |
+
if len(current_chunk) + len(word) > self.chunk_size:
|
70 |
+
chunks.append(current_chunk.strip())
|
71 |
+
current_chunk = word
|
72 |
+
else:
|
73 |
+
current_chunk += word
|
74 |
+
if word[-1] in sentence_endings and len(current_chunk) > self.chunk_size - self.chunk_overlap:
|
75 |
+
chunks.append(current_chunk.strip())
|
76 |
+
current_chunk = ''
|
77 |
+
if current_chunk:
|
78 |
+
chunks.append(current_chunk.strip())
|
79 |
+
if self.chunk_overlap > 0 and len(chunks) > 1:
|
80 |
+
chunks = self._handle_overlap(chunks)
|
81 |
+
return chunks
|
82 |
+
|
83 |
+
def _split_english_text(self, text: str) -> List[str]:
|
84 |
+
# División de texto inglés por frases mediante expresiones regulares
|
85 |
+
sentences = re.split(r'(?<=[.!?])\s+', text.replace('\n', ' '))
|
86 |
+
chunks, current_chunk = [], ''
|
87 |
+
for sentence in sentences:
|
88 |
+
if len(current_chunk) + len(sentence) <= self.chunk_size or not current_chunk:
|
89 |
+
current_chunk += (' ' if current_chunk else '') + sentence
|
90 |
+
else:
|
91 |
+
chunks.append(current_chunk)
|
92 |
+
current_chunk = sentence
|
93 |
+
if current_chunk: # Add the last chunk
|
94 |
+
chunks.append(current_chunk)
|
95 |
+
|
96 |
+
if self.chunk_overlap > 0 and len(chunks) > 1:
|
97 |
+
chunks = self._handle_overlap(chunks)
|
98 |
+
|
99 |
+
return chunks
|
100 |
+
|
101 |
+
def _is_has_chinese(self, text: str) -> bool:
|
102 |
+
# check if contains chinese characters
|
103 |
+
if any("\u4e00" <= ch <= "\u9fff" for ch in text):
|
104 |
+
return True
|
105 |
+
else:
|
106 |
+
return False
|
107 |
+
|
108 |
+
def _handle_overlap(self, chunks: List[str]) -> List[str]:
|
109 |
+
# Tratamiento de los solapamientos entre bloques
|
110 |
+
overlapped_chunks = []
|
111 |
+
for i in range(len(chunks) - 1):
|
112 |
+
chunk = chunks[i] + ' ' + chunks[i + 1][:self.chunk_overlap]
|
113 |
+
overlapped_chunks.append(chunk.strip())
|
114 |
+
overlapped_chunks.append(chunks[-1])
|
115 |
+
return overlapped_chunks
|
116 |
+
|
117 |
+
|
118 |
+
class ChatPDF:
|
119 |
+
def __init__(
|
120 |
+
self,
|
121 |
+
similarity_model: SimilarityABC = None,
|
122 |
+
generate_model_type: str = "auto",
|
123 |
+
generate_model_name_or_path: str = "LenguajeNaturalAI/leniachat-qwen2-1.5B-v0",
|
124 |
+
lora_model_name_or_path: str = None,
|
125 |
+
corpus_files: Union[str, List[str]] = None,
|
126 |
+
save_corpus_emb_dir: str = "corpus_embs/",
|
127 |
+
device: str = None,
|
128 |
+
int8: bool = False,
|
129 |
+
int4: bool = False,
|
130 |
+
chunk_size: int = 250,
|
131 |
+
chunk_overlap: int = 0,
|
132 |
+
rerank_model_name_or_path: str = None,
|
133 |
+
enable_history: bool = False,
|
134 |
+
num_expand_context_chunk: int = 2,
|
135 |
+
similarity_top_k: int = 10,
|
136 |
+
rerank_top_k: int = 3
|
137 |
+
):
|
138 |
+
|
139 |
+
if torch.cuda.is_available():
|
140 |
+
default_device = torch.device(0)
|
141 |
+
elif torch.backends.mps.is_available():
|
142 |
+
default_device = torch.device('cpu')
|
143 |
+
else:
|
144 |
+
default_device = torch.device('cpu')
|
145 |
+
self.device = device or default_device
|
146 |
+
if num_expand_context_chunk > 0 and chunk_overlap > 0:
|
147 |
+
logger.warning(f" 'num_expand_context_chunk' and 'chunk_overlap' cannot both be greater than zero. "
|
148 |
+
f" 'chunk_overlap' has been set to zero by default.")
|
149 |
+
chunk_overlap = 0
|
150 |
+
self.text_splitter = SentenceSplitter(chunk_size, chunk_overlap)
|
151 |
+
if similarity_model is not None:
|
152 |
+
self.sim_model = similarity_model
|
153 |
+
else:
|
154 |
+
m1 = BertSimilarity(model_name_or_path="sentence-transformers/paraphrase-multilingual-MiniLM-L12-v2", device=self.device)
|
155 |
+
m2 = BM25Similarity()
|
156 |
+
default_sim_model = EnsembleSimilarity(similarities=[m1, m2], weights=[0.5, 0.5], c=2)
|
157 |
+
self.sim_model = default_sim_model
|
158 |
+
self.gen_model, self.tokenizer = self._init_gen_model(
|
159 |
+
generate_model_type,
|
160 |
+
generate_model_name_or_path,
|
161 |
+
peft_name=lora_model_name_or_path,
|
162 |
+
int8=int8,
|
163 |
+
int4=int4,
|
164 |
+
)
|
165 |
+
self.history = []
|
166 |
+
self.corpus_files = corpus_files
|
167 |
+
if corpus_files:
|
168 |
+
self.add_corpus(corpus_files)
|
169 |
+
self.save_corpus_emb_dir = save_corpus_emb_dir
|
170 |
+
if rerank_model_name_or_path is None:
|
171 |
+
rerank_model_name_or_path = "maidalun1020/bce-reranker-base_v1"
|
172 |
+
if rerank_model_name_or_path:
|
173 |
+
self.rerank_tokenizer = AutoTokenizer.from_pretrained(rerank_model_name_or_path)
|
174 |
+
self.rerank_model = AutoModelForSequenceClassification.from_pretrained(rerank_model_name_or_path)
|
175 |
+
self.rerank_model.to(self.device)
|
176 |
+
self.rerank_model.eval()
|
177 |
+
else:
|
178 |
+
self.rerank_model = None
|
179 |
+
self.rerank_tokenizer = None
|
180 |
+
self.enable_history = enable_history
|
181 |
+
self.similarity_top_k = similarity_top_k
|
182 |
+
self.num_expand_context_chunk = num_expand_context_chunk
|
183 |
+
self.rerank_top_k = rerank_top_k
|
184 |
+
|
185 |
+
def __str__(self):
|
186 |
+
return f"Similarity model: {self.sim_model}, Generate model: {self.gen_model}"
|
187 |
+
|
188 |
+
def _init_gen_model(
|
189 |
+
self,
|
190 |
+
gen_model_type: str,
|
191 |
+
gen_model_name_or_path: str,
|
192 |
+
peft_name: str = None,
|
193 |
+
int8: bool = False,
|
194 |
+
int4: bool = False,
|
195 |
+
):
|
196 |
+
"""Init generate model."""
|
197 |
+
if int8 or int4:
|
198 |
+
device_map = None
|
199 |
+
else:
|
200 |
+
device_map = "auto"
|
201 |
+
model_class, tokenizer_class = MODEL_CLASSES[gen_model_type]
|
202 |
+
tokenizer = tokenizer_class.from_pretrained(gen_model_name_or_path, trust_remote_code=True)
|
203 |
+
model = model_class.from_pretrained(
|
204 |
+
gen_model_name_or_path,
|
205 |
+
load_in_8bit=int8 if gen_model_type not in ['baichuan', 'chatglm'] else False,
|
206 |
+
load_in_4bit=int4 if gen_model_type not in ['baichuan', 'chatglm'] else False,
|
207 |
+
torch_dtype="auto",
|
208 |
+
device_map=device_map,
|
209 |
+
trust_remote_code=True,
|
210 |
+
)
|
211 |
+
if self.device == torch.device('cpu'):
|
212 |
+
model.float()
|
213 |
+
if gen_model_type in ['baichuan', 'chatglm']:
|
214 |
+
if int4:
|
215 |
+
model = model.quantize(4).cuda()
|
216 |
+
elif int8:
|
217 |
+
model = model.quantize(8).cuda()
|
218 |
+
try:
|
219 |
+
model.generation_config = GenerationConfig.from_pretrained(gen_model_name_or_path, trust_remote_code=True)
|
220 |
+
except Exception as e:
|
221 |
+
logger.warning(f"No se pudo cargar la configuración de generación desde {gen_model_name_or_path}, {e}")
|
222 |
+
if peft_name:
|
223 |
+
model = PeftModel.from_pretrained(
|
224 |
+
model,
|
225 |
+
peft_name,
|
226 |
+
torch_dtype="auto",
|
227 |
+
)
|
228 |
+
logger.info(f"Modelo peft cargado desde {peft_name}")
|
229 |
+
model.eval()
|
230 |
+
return model, tokenizer
|
231 |
+
|
232 |
+
def _get_chat_input(self):
|
233 |
+
messages = []
|
234 |
+
for conv in self.history:
|
235 |
+
if conv and len(conv) > 0 and conv[0]:
|
236 |
+
messages.append({'role': 'user', 'content': conv[0]})
|
237 |
+
if conv and len(conv) > 1 and conv[1]:
|
238 |
+
messages.append({'role': 'assistant', 'content': conv[1]})
|
239 |
+
input_ids = self.tokenizer.apply_chat_template(
|
240 |
+
conversation=messages,
|
241 |
+
tokenize=True,
|
242 |
+
add_generation_prompt=True,
|
243 |
+
return_tensors='pt'
|
244 |
+
)
|
245 |
+
return input_ids.to(self.gen_model.device)
|
246 |
+
|
247 |
+
@torch.inference_mode()
|
248 |
+
def stream_generate_answer(
|
249 |
+
self,
|
250 |
+
max_new_tokens=512,
|
251 |
+
temperature=0.7,
|
252 |
+
repetition_penalty=1.0,
|
253 |
+
context_len=2048
|
254 |
+
):
|
255 |
+
streamer = TextIteratorStreamer(self.tokenizer, timeout=520.0, skip_prompt=True, skip_special_tokens=True)
|
256 |
+
input_ids = self._get_chat_input()
|
257 |
+
max_src_len = context_len - max_new_tokens - 8
|
258 |
+
input_ids = input_ids[-max_src_len:]
|
259 |
+
generation_kwargs = dict(
|
260 |
+
input_ids=input_ids,
|
261 |
+
max_new_tokens=max_new_tokens,
|
262 |
+
temperature=temperature,
|
263 |
+
do_sample=True,
|
264 |
+
repetition_penalty=repetition_penalty,
|
265 |
+
streamer=streamer,
|
266 |
+
)
|
267 |
+
thread = Thread(target=self.gen_model.generate, kwargs=generation_kwargs)
|
268 |
+
thread.start()
|
269 |
+
|
270 |
+
yield from streamer
|
271 |
+
|
272 |
+
def add_corpus(self, files: Union[str, List[str]]):
|
273 |
+
"""Load document files."""
|
274 |
+
if isinstance(files, str):
|
275 |
+
files = [files]
|
276 |
+
for doc_file in files:
|
277 |
+
if doc_file.endswith('.pdf'):
|
278 |
+
corpus = self.extract_text_from_pdf(doc_file)
|
279 |
+
elif doc_file.endswith('.docx'):
|
280 |
+
corpus = self.extract_text_from_docx(doc_file)
|
281 |
+
elif doc_file.endswith('.md'):
|
282 |
+
corpus = self.extract_text_from_markdown(doc_file)
|
283 |
+
else:
|
284 |
+
corpus = self.extract_text_from_txt(doc_file)
|
285 |
+
full_text = '\n'.join(corpus)
|
286 |
+
chunks = self.text_splitter.split_text(full_text)
|
287 |
+
self.sim_model.add_corpus(chunks)
|
288 |
+
self.corpus_files = files
|
289 |
+
logger.debug(f"files: {files}, corpus size: {len(self.sim_model.corpus)}, top3: "
|
290 |
+
f"{list(self.sim_model.corpus.values())[:3]}")
|
291 |
+
|
292 |
+
@staticmethod
|
293 |
+
def get_file_hash(fpaths):
|
294 |
+
hasher = hashlib.md5()
|
295 |
+
target_file_data = bytes()
|
296 |
+
if isinstance(fpaths, str):
|
297 |
+
fpaths = [fpaths]
|
298 |
+
for fpath in fpaths:
|
299 |
+
with open(fpath, 'rb') as file:
|
300 |
+
chunk = file.read(1024 * 1024) # read only first 1MB
|
301 |
+
hasher.update(chunk)
|
302 |
+
target_file_data += chunk
|
303 |
+
|
304 |
+
hash_name = hasher.hexdigest()[:32]
|
305 |
+
return hash_name
|
306 |
+
|
307 |
+
@staticmethod
|
308 |
+
def extract_text_from_pdf(file_path: str):
|
309 |
+
"""Extract text content from a PDF file."""
|
310 |
+
import PyPDF2
|
311 |
+
contents = []
|
312 |
+
with open(file_path, 'rb') as f:
|
313 |
+
pdf_reader = PyPDF2.PdfReader(f)
|
314 |
+
for page in pdf_reader.pages:
|
315 |
+
page_text = page.extract_text().strip()
|
316 |
+
raw_text = [text.strip() for text in page_text.splitlines() if text.strip()]
|
317 |
+
new_text = ''
|
318 |
+
for text in raw_text:
|
319 |
+
# Añadir un espacio antes de concatenar si new_text no está vacío
|
320 |
+
if new_text:
|
321 |
+
new_text += ' '
|
322 |
+
new_text += text
|
323 |
+
if text[-1] in ['.', '!', '?', '。', '!', '?', '…', ';', ';', ':', ':', '”', '’', ')', '】', '》', '」',
|
324 |
+
'』', '〕', '〉', '》', '〗', '〞', '〟', '»', '"', "'", ')', ']', '}']:
|
325 |
+
contents.append(new_text)
|
326 |
+
new_text = ''
|
327 |
+
if new_text:
|
328 |
+
contents.append(new_text)
|
329 |
+
return contents
|
330 |
+
|
331 |
+
@staticmethod
|
332 |
+
def extract_text_from_txt(file_path: str):
|
333 |
+
"""Extract text content from a TXT file."""
|
334 |
+
with open(file_path, 'r', encoding='utf-8') as f:
|
335 |
+
contents = [text.strip() for text in f.readlines() if text.strip()]
|
336 |
+
return contents
|
337 |
+
|
338 |
+
@staticmethod
|
339 |
+
def extract_text_from_docx(file_path: str):
|
340 |
+
"""Extract text content from a DOCX file."""
|
341 |
+
import docx
|
342 |
+
document = docx.Document(file_path)
|
343 |
+
contents = [paragraph.text.strip() for paragraph in document.paragraphs if paragraph.text.strip()]
|
344 |
+
return contents
|
345 |
+
|
346 |
+
@staticmethod
|
347 |
+
def extract_text_from_markdown(file_path: str):
|
348 |
+
"""Extract text content from a Markdown file."""
|
349 |
+
import markdown
|
350 |
+
from bs4 import BeautifulSoup
|
351 |
+
with open(file_path, 'r', encoding='utf-8') as f:
|
352 |
+
markdown_text = f.read()
|
353 |
+
html = markdown.markdown(markdown_text)
|
354 |
+
soup = BeautifulSoup(html, 'html.parser')
|
355 |
+
contents = [text.strip() for text in soup.get_text().splitlines() if text.strip()]
|
356 |
+
return contents
|
357 |
+
|
358 |
+
@staticmethod
|
359 |
+
def _add_source_numbers(lst):
|
360 |
+
"""Add source numbers to a list of strings."""
|
361 |
+
return [f'[{idx + 1}]\t "{item}"' for idx, item in enumerate(lst)]
|
362 |
+
|
363 |
+
def _get_reranker_score(self, query: str, reference_results: List[str]):
|
364 |
+
"""Get reranker score."""
|
365 |
+
pairs = []
|
366 |
+
for reference in reference_results:
|
367 |
+
pairs.append([query, reference])
|
368 |
+
with torch.no_grad():
|
369 |
+
inputs = self.rerank_tokenizer(pairs, padding=True, truncation=True, return_tensors='pt', max_length=512)
|
370 |
+
inputs_on_device = {k: v.to(self.rerank_model.device) for k, v in inputs.items()}
|
371 |
+
scores = self.rerank_model(**inputs_on_device, return_dict=True).logits.view(-1, ).float()
|
372 |
+
|
373 |
+
return scores
|
374 |
+
|
375 |
+
def get_reference_results(self, query: str):
|
376 |
+
"""
|
377 |
+
Get reference results.
|
378 |
+
1. Similarity model get similar chunks
|
379 |
+
2. Rerank similar chunks
|
380 |
+
3. Expand reference context chunk
|
381 |
+
:param query:
|
382 |
+
:return:
|
383 |
+
"""
|
384 |
+
reference_results = []
|
385 |
+
sim_contents = self.sim_model.most_similar(query, topn=self.similarity_top_k)
|
386 |
+
# Get reference results from corpus
|
387 |
+
hit_chunk_dict = dict()
|
388 |
+
for query_id, id_score_dict in sim_contents.items():
|
389 |
+
for corpus_id, s in id_score_dict.items():
|
390 |
+
hit_chunk = self.sim_model.corpus[corpus_id]
|
391 |
+
reference_results.append(hit_chunk)
|
392 |
+
hit_chunk_dict[corpus_id] = hit_chunk
|
393 |
+
|
394 |
+
if reference_results:
|
395 |
+
if self.rerank_model is not None:
|
396 |
+
# Rerank reference results
|
397 |
+
rerank_scores = self._get_reranker_score(query, reference_results)
|
398 |
+
logger.debug(f"rerank_scores: {rerank_scores}")
|
399 |
+
# Get rerank top k chunks
|
400 |
+
reference_results = [reference for reference, score in sorted(
|
401 |
+
zip(reference_results, rerank_scores), key=lambda x: x[1], reverse=True)][:self.rerank_top_k]
|
402 |
+
hit_chunk_dict = {corpus_id: hit_chunk for corpus_id, hit_chunk in hit_chunk_dict.items() if
|
403 |
+
hit_chunk in reference_results}
|
404 |
+
# Expand reference context chunk
|
405 |
+
if self.num_expand_context_chunk > 0:
|
406 |
+
new_reference_results = []
|
407 |
+
for corpus_id, hit_chunk in hit_chunk_dict.items():
|
408 |
+
expanded_reference = self.sim_model.corpus.get(corpus_id - 1, '') + hit_chunk
|
409 |
+
for i in range(self.num_expand_context_chunk):
|
410 |
+
expanded_reference += self.sim_model.corpus.get(corpus_id + i + 1, '')
|
411 |
+
new_reference_results.append(expanded_reference)
|
412 |
+
reference_results = new_reference_results
|
413 |
+
return reference_results
|
414 |
+
|
415 |
+
def predict_stream(
|
416 |
+
self,
|
417 |
+
query: str,
|
418 |
+
max_length: int = 512,
|
419 |
+
context_len: int = 2048,
|
420 |
+
temperature: float = 0.7,
|
421 |
+
):
|
422 |
+
"""Generate predictions stream."""
|
423 |
+
stop_str = self.tokenizer.eos_token if self.tokenizer.eos_token else "</s>"
|
424 |
+
if not self.enable_history:
|
425 |
+
self.history = []
|
426 |
+
if self.sim_model.corpus:
|
427 |
+
reference_results = self.get_reference_results(query)
|
428 |
+
if not reference_results:
|
429 |
+
yield 'No se ha proporcionado suficiente información relevante', reference_results
|
430 |
+
reference_results = self._add_source_numbers(reference_results)
|
431 |
+
context_str = '\n'.join(reference_results)[:]
|
432 |
+
#print("context_str: " , (context_len - len(PROMPT_TEMPLATE)))
|
433 |
+
prompt = PROMPT_TEMPLATE.format(context_str=context_str, query_str=query)
|
434 |
+
logger.debug(f"prompt: {prompt}")
|
435 |
+
else:
|
436 |
+
prompt = query
|
437 |
+
logger.debug(prompt)
|
438 |
+
self.history.append([prompt, ''])
|
439 |
+
response = ""
|
440 |
+
for new_text in self.stream_generate_answer(
|
441 |
+
max_new_tokens=max_length,
|
442 |
+
temperature=temperature,
|
443 |
+
context_len=context_len,
|
444 |
+
):
|
445 |
+
if new_text != stop_str:
|
446 |
+
response += new_text
|
447 |
+
yield response
|
448 |
+
|
449 |
+
def predict(
|
450 |
+
self,
|
451 |
+
query: str,
|
452 |
+
max_length: int = 512,
|
453 |
+
context_len: int = 2048,
|
454 |
+
temperature: float = 0.7,
|
455 |
+
):
|
456 |
+
"""Query from corpus."""
|
457 |
+
reference_results = []
|
458 |
+
if not self.enable_history:
|
459 |
+
self.history = []
|
460 |
+
if self.sim_model.corpus:
|
461 |
+
reference_results = self.get_reference_results(query)
|
462 |
+
|
463 |
+
if not reference_results:
|
464 |
+
return 'No se ha proporcionado suficiente información relevante', reference_results
|
465 |
+
reference_results = self._add_source_numbers(reference_results)
|
466 |
+
#context_str = '\n'.join(reference_results) # Usa todos los fragmentos
|
467 |
+
context_st = '\n'.join(reference_results)[:(context_len - len(PROMPT_TEMPLATE))]
|
468 |
+
#print("Context: ", (context_len - len(PROMPT_TEMPLATE)))
|
469 |
+
print(".......................................................")
|
470 |
+
context_str = '\n'.join(reference_results)[:]
|
471 |
+
#print("context_str: ", context_str)
|
472 |
+
prompt = PROMPT_TEMPLATE.format(context_str=context_str, query_str=query)
|
473 |
+
logger.debug(f"prompt: {prompt}")
|
474 |
+
else:
|
475 |
+
prompt = query
|
476 |
+
self.history.append([prompt, ''])
|
477 |
+
response = ""
|
478 |
+
for new_text in self.stream_generate_answer(
|
479 |
+
max_new_tokens=max_length,
|
480 |
+
temperature=temperature,
|
481 |
+
context_len=context_len,
|
482 |
+
):
|
483 |
+
response += new_text
|
484 |
+
response = response.strip()
|
485 |
+
self.history[-1][1] = response
|
486 |
+
return response, reference_results
|
487 |
+
|
488 |
+
def save_corpus_emb(self):
|
489 |
+
dir_name = self.get_file_hash(self.corpus_files)
|
490 |
+
save_dir = os.path.join(self.save_corpus_emb_dir, dir_name)
|
491 |
+
if hasattr(self.sim_model, 'save_corpus_embeddings'):
|
492 |
+
self.sim_model.save_corpus_embeddings(save_dir)
|
493 |
+
logger.debug(f"Saving corpus embeddings to {save_dir}")
|
494 |
+
return save_dir
|
495 |
+
|
496 |
+
def load_corpus_emb(self, emb_dir: str):
|
497 |
+
if hasattr(self.sim_model, 'load_corpus_embeddings'):
|
498 |
+
logger.debug(f"Loading corpus embeddings from {emb_dir}")
|
499 |
+
self.sim_model.load_corpus_embeddings(emb_dir)
|
500 |
+
|
501 |
+
def save_corpus_text(self):
|
502 |
+
if not self.corpus_files:
|
503 |
+
logger.warning("No hay archivos de corpus para guardar.")
|
504 |
+
return
|
505 |
+
|
506 |
+
corpus_text_file = os.path.join("corpus_embs/", "corpus_text.txt")
|
507 |
+
|
508 |
+
with open(corpus_text_file, 'w', encoding='utf-8') as f:
|
509 |
+
for chunk in self.sim_model.corpus.values():
|
510 |
+
f.write(chunk + "\n\n") # Añade dos saltos de línea entre chunks para mejor legibilidad
|
511 |
+
|
512 |
+
logger.info(f"Texto del corpus guardado en: {corpus_text_file}")
|
513 |
+
return corpus_text_file
|
514 |
+
|
515 |
+
def load_corpus_text(self, emb_dir: str):
|
516 |
+
corpus_text_file = os.path.join("corpus_embs/", "corpus_text.txt")
|
517 |
+
if os.path.exists(corpus_text_file):
|
518 |
+
with open(corpus_text_file, 'r', encoding='utf-8') as f:
|
519 |
+
corpus_text = f.read().split("\n\n") # Asumiendo que usamos dos saltos de línea como separador
|
520 |
+
self.sim_model.corpus = {i: chunk.strip() for i, chunk in enumerate(corpus_text) if chunk.strip()}
|
521 |
+
logger.info(f"Texto del corpus cargado desde: {corpus_text_file}")
|
522 |
+
else:
|
523 |
+
logger.warning(f"No se encontró el archivo de texto del corpus en: {corpus_text_file}")
|
524 |
+
|
525 |
+
if __name__ == "__main__":
|
526 |
+
parser = argparse.ArgumentParser()
|
527 |
+
parser.add_argument("--sim_model_name", type=str, default="sentence-transformers/paraphrase-multilingual-MiniLM-L12-v2")
|
528 |
+
parser.add_argument("--gen_model_type", type=str, default="auto")
|
529 |
+
parser.add_argument("--gen_model_name", type=str, default="LenguajeNaturalAI/leniachat-qwen2-1.5B-v0")
|
530 |
+
parser.add_argument("--lora_model", type=str, default=None)
|
531 |
+
parser.add_argument("--rerank_model_name", type=str, default="maidalun1020/bce-reranker-base_v1")
|
532 |
+
parser.add_argument("--corpus_files", type=str, default="docs/corpus.txt")
|
533 |
+
parser.add_argument("--device", type=str, default=None)
|
534 |
+
parser.add_argument("--int4", action='store_true', help="use int4 quantization")
|
535 |
+
parser.add_argument("--int8", action='store_true', help="use int8 quantization")
|
536 |
+
parser.add_argument("--chunk_size", type=int, default=220)
|
537 |
+
parser.add_argument("--chunk_overlap", type=int, default=50)
|
538 |
+
parser.add_argument("--num_expand_context_chunk", type=int, default=2)
|
539 |
+
args = parser.parse_args()
|
540 |
+
print(args)
|
541 |
+
sim_model = BertSimilarity(model_name_or_path=args.sim_model_name, device=args.device)
|
542 |
+
m = ChatPDF(
|
543 |
+
similarity_model=sim_model,
|
544 |
+
generate_model_type=args.gen_model_type,
|
545 |
+
generate_model_name_or_path=args.gen_model_name,
|
546 |
+
lora_model_name_or_path=args.lora_model,
|
547 |
+
device=args.device,
|
548 |
+
int4=args.int4,
|
549 |
+
int8=args.int8,
|
550 |
+
chunk_size=args.chunk_size,
|
551 |
+
chunk_overlap=args.chunk_overlap,
|
552 |
+
corpus_files=args.corpus_files.split(','),
|
553 |
+
num_expand_context_chunk=args.num_expand_context_chunk,
|
554 |
+
rerank_model_name_or_path=args.rerank_model_name,
|
555 |
+
)
|
556 |
+
logger.info(f"chatpdf model: {m}")
|
557 |
+
|
558 |
+
# Comprobar si existen incrustaciones guardadas
|
559 |
+
dir_name = m.get_file_hash(args.corpus_files.split(','))
|
560 |
+
save_dir = os.path.join(m.save_corpus_emb_dir, dir_name)
|
561 |
+
|
562 |
+
if os.path.exists(save_dir):
|
563 |
+
# Cargar las incrustaciones guardadas
|
564 |
+
m.load_corpus_emb(save_dir)
|
565 |
+
print(f"Incrustaciones del corpus cargadas desde: {save_dir}")
|
566 |
+
else:
|
567 |
+
# Procesar el corpus y guardar las incrustaciones
|
568 |
+
m.add_corpus(args.corpus_files.split(','))
|
569 |
+
save_dir = m.save_corpus_emb()
|
570 |
+
# Guardar el texto del corpus
|
571 |
+
m.save_corpus_text()
|
572 |
+
print(f"Las incrustaciones del corpus se han guardado en: {save_dir}")
|
573 |
+
|
574 |
+
while True:
|
575 |
+
query = input("\nEnter a query: ")
|
576 |
+
if query == "exit":
|
577 |
+
break
|
578 |
+
if query.strip() == "":
|
579 |
+
continue
|
580 |
+
r, refs = m.predict(query)
|
581 |
+
print(r, refs)
|
582 |
+
print("\nRespuesta: ", r)
|