Spaces:
Running
Running
Update app.py
Browse files
app.py
CHANGED
|
@@ -1,9 +1,13 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
import torch
|
|
|
|
| 2 |
import gradio as gr
|
| 3 |
import spaces
|
| 4 |
from transformers import AutoModelForCausalLM, AutoTokenizer, TextIteratorStreamer
|
| 5 |
|
| 6 |
-
import os
|
| 7 |
from threading import Thread
|
| 8 |
import random
|
| 9 |
from datasets import load_dataset
|
|
@@ -17,15 +21,29 @@ import pyarrow.parquet as pq
|
|
| 17 |
import pypdf
|
| 18 |
import io
|
| 19 |
import pyarrow.parquet as pq
|
| 20 |
-
from
|
| 21 |
-
from pdfminer.layout import LAParams
|
| 22 |
-
from tabulate import tabulate # tabulate ์ถ๊ฐ
|
| 23 |
import platform
|
| 24 |
import subprocess
|
| 25 |
import pytesseract
|
| 26 |
from pdf2image import convert_from_path
|
| 27 |
|
| 28 |
-
#
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 29 |
current_file_context = None
|
| 30 |
|
| 31 |
# ํ๊ฒฝ ๋ณ์ ์ค์
|
|
@@ -48,6 +66,7 @@ vectorizer = TfidfVectorizer(max_features=1000)
|
|
| 48 |
question_vectors = vectorizer.fit_transform(questions)
|
| 49 |
print("TF-IDF ๋ฒกํฐํ ์๋ฃ")
|
| 50 |
|
|
|
|
| 51 |
class ChatHistory:
|
| 52 |
def __init__(self):
|
| 53 |
self.history = []
|
|
@@ -103,19 +122,18 @@ class ChatHistory:
|
|
| 103 |
print(f"ํ์คํ ๋ฆฌ ๋ก๋ ์คํจ: {e}")
|
| 104 |
self.history = []
|
| 105 |
|
|
|
|
| 106 |
# ์ ์ญ ChatHistory ์ธ์คํด์ค ์์ฑ
|
| 107 |
chat_history = ChatHistory()
|
| 108 |
|
|
|
|
| 109 |
def find_relevant_context(query, top_k=3):
|
| 110 |
# ์ฟผ๋ฆฌ ๋ฒกํฐํ
|
| 111 |
query_vector = vectorizer.transform([query])
|
| 112 |
-
|
| 113 |
# ์ฝ์ฌ์ธ ์ ์ฌ๋ ๊ณ์ฐ
|
| 114 |
similarities = (query_vector * question_vectors.T).toarray()[0]
|
| 115 |
-
|
| 116 |
# ๊ฐ์ฅ ์ ์ฌํ ์ง๋ฌธ๋ค์ ์ธ๋ฑ์ค
|
| 117 |
top_indices = np.argsort(similarities)[-top_k:][::-1]
|
| 118 |
-
|
| 119 |
# ๊ด๋ จ ์ปจํ
์คํธ ์ถ์ถ
|
| 120 |
relevant_contexts = []
|
| 121 |
for idx in top_indices:
|
|
@@ -125,14 +143,74 @@ def find_relevant_context(query, top_k=3):
|
|
| 125 |
'answer': wiki_dataset['train']['answer'][idx],
|
| 126 |
'similarity': similarities[idx]
|
| 127 |
})
|
| 128 |
-
|
| 129 |
return relevant_contexts
|
| 130 |
|
|
|
|
| 131 |
def init_msg():
|
| 132 |
-
return "
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 133 |
|
| 134 |
def analyze_file_content(content, file_type):
|
| 135 |
-
"""
|
| 136 |
if file_type in ['parquet', 'csv']:
|
| 137 |
try:
|
| 138 |
lines = content.split('\n')
|
|
@@ -142,115 +220,87 @@ def analyze_file_content(content, file_type):
|
|
| 142 |
return f"๐ Dataset Structure: {columns} columns, {rows} rows"
|
| 143 |
except:
|
| 144 |
return "โ Failed to analyze dataset structure"
|
| 145 |
-
|
| 146 |
lines = content.split('\n')
|
| 147 |
total_lines = len(lines)
|
| 148 |
non_empty_lines = len([line for line in lines if line.strip()])
|
| 149 |
-
|
| 150 |
if any(keyword in content.lower() for keyword in ['def ', 'class ', 'import ', 'function']):
|
| 151 |
functions = len([line for line in lines if 'def ' in line])
|
| 152 |
classes = len([line for line in lines if 'class ' in line])
|
| 153 |
imports = len([line for line in lines if 'import ' in line or 'from ' in line])
|
| 154 |
return f"๐ป Code Structure: {total_lines} lines (Functions: {functions}, Classes: {classes}, Imports: {imports})"
|
| 155 |
-
|
| 156 |
paragraphs = content.count('\n\n') + 1
|
| 157 |
words = len(content.split())
|
| 158 |
return f"๐ Document Structure: {total_lines} lines, {paragraphs} paragraphs, approximately {words} words"
|
| 159 |
|
|
|
|
| 160 |
def read_uploaded_file(file):
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 161 |
if file is None:
|
| 162 |
return "", ""
|
| 163 |
try:
|
| 164 |
file_ext = os.path.splitext(file.name)[1].lower()
|
| 165 |
-
|
| 166 |
-
# Parquet
|
| 167 |
if file_ext == '.parquet':
|
| 168 |
try:
|
| 169 |
table = pq.read_table(file.name)
|
| 170 |
df = table.to_pandas()
|
| 171 |
-
|
| 172 |
content = f"๐ Parquet File Analysis:\n\n"
|
| 173 |
content += f"1. Basic Information:\n"
|
| 174 |
content += f"- Total Rows: {len(df):,}\n"
|
| 175 |
content += f"- Total Columns: {len(df.columns)}\n"
|
| 176 |
content += f"- Memory Usage: {df.memory_usage(deep=True).sum() / 1024 / 1024:.2f} MB\n\n"
|
| 177 |
-
|
| 178 |
content += f"2. Column Information:\n"
|
| 179 |
for col in df.columns:
|
| 180 |
content += f"- {col} ({df[col].dtype})\n"
|
| 181 |
-
|
| 182 |
content += f"\n3. Data Preview:\n"
|
| 183 |
content += tabulate(df.head(5), headers='keys', tablefmt='pipe', showindex=False)
|
| 184 |
-
|
| 185 |
content += f"\n\n4. Missing Values:\n"
|
| 186 |
null_counts = df.isnull().sum()
|
| 187 |
for col, count in null_counts[null_counts > 0].items():
|
| 188 |
content += f"- {col}: {count:,} ({count/len(df)*100:.1f}%)\n"
|
| 189 |
-
|
| 190 |
numeric_cols = df.select_dtypes(include=['int64', 'float64']).columns
|
| 191 |
if len(numeric_cols) > 0:
|
| 192 |
content += f"\n5. Numeric Column Statistics:\n"
|
| 193 |
stats_df = df[numeric_cols].describe()
|
| 194 |
content += tabulate(stats_df, headers='keys', tablefmt='pipe')
|
| 195 |
-
|
| 196 |
return content, "parquet"
|
| 197 |
except Exception as e:
|
| 198 |
return f"Error reading Parquet file: {str(e)}", "error"
|
| 199 |
-
|
| 200 |
-
# PDF
|
| 201 |
if file_ext == '.pdf':
|
| 202 |
try:
|
| 203 |
-
|
| 204 |
-
|
| 205 |
-
|
| 206 |
-
|
| 207 |
-
content
|
| 208 |
-
content +=
|
| 209 |
-
|
| 210 |
-
|
| 211 |
-
|
| 212 |
-
|
| 213 |
-
|
| 214 |
-
|
| 215 |
-
|
| 216 |
-
try:
|
| 217 |
-
text = extract_text(
|
| 218 |
-
file.name,
|
| 219 |
-
laparams=LAParams(
|
| 220 |
-
line_margin=0.5,
|
| 221 |
-
word_margin=0.1,
|
| 222 |
-
char_margin=2.0,
|
| 223 |
-
all_texts=True
|
| 224 |
-
)
|
| 225 |
-
)
|
| 226 |
-
except:
|
| 227 |
-
text = ""
|
| 228 |
-
|
| 229 |
-
if not text.strip():
|
| 230 |
-
text = extract_pdf_text_with_ocr(file.name)
|
| 231 |
-
|
| 232 |
-
if text:
|
| 233 |
-
words = text.split()
|
| 234 |
-
lines = text.split('\n')
|
| 235 |
-
content += f"\n3. Text Analysis:\n"
|
| 236 |
-
content += f"- Total Words: {len(words):,}\n"
|
| 237 |
-
content += f"- Unique Words: {len(set(words)):,}\n"
|
| 238 |
-
content += f"- Total Lines: {len(lines):,}\n"
|
| 239 |
-
|
| 240 |
-
content += f"\n4. Content Preview:\n"
|
| 241 |
-
preview_length = min(2000, len(text))
|
| 242 |
-
content += f"--- First {preview_length} characters ---\n"
|
| 243 |
-
content += text[:preview_length]
|
| 244 |
-
if len(text) > preview_length:
|
| 245 |
-
content += f"\n... (Showing partial content of {len(text):,} characters)\n"
|
| 246 |
-
else:
|
| 247 |
-
content += "\nโ ๏ธ Text extraction failed"
|
| 248 |
-
|
| 249 |
return content, "pdf"
|
| 250 |
except Exception as e:
|
| 251 |
return f"Error reading PDF file: {str(e)}", "error"
|
| 252 |
-
|
| 253 |
-
# CSV
|
| 254 |
elif file_ext == '.csv':
|
| 255 |
encodings = ['utf-8', 'cp949', 'euc-kr', 'latin1']
|
| 256 |
for encoding in encodings:
|
|
@@ -261,44 +311,44 @@ def read_uploaded_file(file):
|
|
| 261 |
content += f"- Total Rows: {len(df):,}\n"
|
| 262 |
content += f"- Total Columns: {len(df.columns)}\n"
|
| 263 |
content += f"- Memory Usage: {df.memory_usage(deep=True).sum() / 1024 / 1024:.2f} MB\n\n"
|
| 264 |
-
|
| 265 |
content += f"2. Column Information:\n"
|
| 266 |
for col in df.columns:
|
| 267 |
content += f"- {col} ({df[col].dtype})\n"
|
| 268 |
-
|
| 269 |
content += f"\n3. Data Preview:\n"
|
| 270 |
content += df.head(5).to_markdown(index=False)
|
| 271 |
-
|
| 272 |
content += f"\n\n4. Missing Values:\n"
|
| 273 |
null_counts = df.isnull().sum()
|
| 274 |
for col, count in null_counts[null_counts > 0].items():
|
| 275 |
content += f"- {col}: {count:,} ({count/len(df)*100:.1f}%)\n"
|
| 276 |
-
|
| 277 |
return content, "csv"
|
| 278 |
except UnicodeDecodeError:
|
| 279 |
continue
|
| 280 |
raise UnicodeDecodeError(f"Unable to read file with supported encodings ({', '.join(encodings)})")
|
| 281 |
-
|
| 282 |
-
#
|
| 283 |
else:
|
| 284 |
encodings = ['utf-8', 'cp949', 'euc-kr', 'latin1']
|
| 285 |
for encoding in encodings:
|
| 286 |
try:
|
| 287 |
with open(file.name, 'r', encoding=encoding) as f:
|
| 288 |
content = f.read()
|
| 289 |
-
|
| 290 |
lines = content.split('\n')
|
| 291 |
total_lines = len(lines)
|
| 292 |
non_empty_lines = len([line for line in lines if line.strip()])
|
| 293 |
-
|
| 294 |
is_code = any(keyword in content.lower() for keyword in ['def ', 'class ', 'import ', 'function'])
|
| 295 |
-
|
| 296 |
analysis = f"\n๐ File Analysis:\n"
|
| 297 |
if is_code:
|
| 298 |
functions = len([line for line in lines if 'def ' in line])
|
| 299 |
classes = len([line for line in lines if 'class ' in line])
|
| 300 |
imports = len([line for line in lines if 'import ' in line or 'from ' in line])
|
| 301 |
-
|
| 302 |
analysis += f"- File Type: Code\n"
|
| 303 |
analysis += f"- Total Lines: {total_lines:,}\n"
|
| 304 |
analysis += f"- Functions: {functions}\n"
|
|
@@ -307,18 +357,18 @@ def read_uploaded_file(file):
|
|
| 307 |
else:
|
| 308 |
words = len(content.split())
|
| 309 |
chars = len(content)
|
| 310 |
-
|
| 311 |
analysis += f"- File Type: Text\n"
|
| 312 |
analysis += f"- Total Lines: {total_lines:,}\n"
|
| 313 |
analysis += f"- Non-empty Lines: {non_empty_lines:,}\n"
|
| 314 |
analysis += f"- Word Count: {words:,}\n"
|
| 315 |
analysis += f"- Character Count: {chars:,}\n"
|
| 316 |
-
|
| 317 |
return content + analysis, "text"
|
| 318 |
except UnicodeDecodeError:
|
| 319 |
continue
|
| 320 |
raise UnicodeDecodeError(f"Unable to read file with supported encodings ({', '.join(encodings)})")
|
| 321 |
-
|
| 322 |
except Exception as e:
|
| 323 |
return f"Error reading file: {str(e)}", "error"
|
| 324 |
|
|
@@ -461,10 +511,10 @@ body {
|
|
| 461 |
font-size: 1.1em !important;
|
| 462 |
padding: 10px 15px !important;
|
| 463 |
display: flex !important;
|
| 464 |
-
align-items: flex-start !important;
|
| 465 |
}
|
| 466 |
.input-textbox textarea {
|
| 467 |
-
padding-top: 5px !important;
|
| 468 |
}
|
| 469 |
.send-button {
|
| 470 |
height: 70px !important;
|
|
@@ -478,73 +528,118 @@ body {
|
|
| 478 |
}
|
| 479 |
"""
|
| 480 |
|
| 481 |
-
# GPU ๋ฉ๋ชจ๋ฆฌ ๊ด๋ฆฌ ํจ์ ์์
|
| 482 |
def clear_cuda_memory():
|
| 483 |
if hasattr(torch.cuda, 'empty_cache'):
|
| 484 |
with torch.cuda.device('cuda'):
|
| 485 |
torch.cuda.empty_cache()
|
| 486 |
|
| 487 |
-
|
| 488 |
@spaces.GPU
|
| 489 |
def load_model():
|
| 490 |
try:
|
| 491 |
-
|
| 492 |
MODEL_ID,
|
| 493 |
torch_dtype=torch.bfloat16,
|
| 494 |
device_map="auto",
|
| 495 |
)
|
| 496 |
-
return
|
| 497 |
except Exception as e:
|
| 498 |
print(f"๋ชจ๋ธ ๋ก๋ ์ค๋ฅ: {str(e)}")
|
| 499 |
raise
|
| 500 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 501 |
@spaces.GPU
|
| 502 |
-
def stream_chat(
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 503 |
global model, current_file_context
|
| 504 |
-
|
| 505 |
try:
|
| 506 |
if model is None:
|
| 507 |
model = load_model()
|
| 508 |
-
|
| 509 |
print(f'message is - {message}')
|
| 510 |
print(f'history is - {history}')
|
| 511 |
|
| 512 |
# ํ์ผ ์
๋ก๋ ์ฒ๋ฆฌ
|
| 513 |
file_context = ""
|
| 514 |
if uploaded_file and message == "ํ์ผ์ ๋ถ์ํ๊ณ ์์ต๋๋ค...":
|
|
|
|
|
|
|
| 515 |
try:
|
| 516 |
content, file_type = read_uploaded_file(uploaded_file)
|
| 517 |
if content:
|
| 518 |
file_analysis = analyze_file_content(content, file_type)
|
| 519 |
-
file_context =
|
|
|
|
|
|
|
|
|
|
| 520 |
current_file_context = file_context # ํ์ผ ์ปจํ
์คํธ ์ ์ฅ
|
| 521 |
message = "์
๋ก๋๋ ํ์ผ์ ๋ถ์ํด์ฃผ์ธ์."
|
| 522 |
except Exception as e:
|
| 523 |
print(f"ํ์ผ ๋ถ์ ์ค๋ฅ: {str(e)}")
|
| 524 |
file_context = f"\n\nโ ํ์ผ ๋ถ์ ์ค ์ค๋ฅ๊ฐ ๋ฐ์ํ์ต๋๋ค: {str(e)}"
|
| 525 |
-
elif current_file_context:
|
|
|
|
| 526 |
file_context = current_file_context
|
| 527 |
-
|
| 528 |
|
| 529 |
# ๋ฉ๋ชจ๋ฆฌ ์ฌ์ฉ๋ ๋ชจ๋ํฐ๋ง
|
| 530 |
if torch.cuda.is_available():
|
| 531 |
print(f"CUDA ๋ฉ๋ชจ๋ฆฌ ์ฌ์ฉ๋: {torch.cuda.memory_allocated() / 1024**2:.2f} MB")
|
| 532 |
|
| 533 |
# ๋ํ ํ์คํ ๋ฆฌ๊ฐ ๋๋ฌด ๊ธธ๋ฉด ์๋ผ๋ด๊ธฐ
|
| 534 |
-
max_history_length = 10
|
| 535 |
if len(history) > max_history_length:
|
| 536 |
history = history[-max_history_length:]
|
| 537 |
|
| 538 |
-
#
|
| 539 |
try:
|
| 540 |
relevant_contexts = find_relevant_context(message)
|
| 541 |
wiki_context = "\n\n๊ด๋ จ ์ํคํผ๋์ ์ ๋ณด:\n"
|
| 542 |
for ctx in relevant_contexts:
|
| 543 |
-
wiki_context +=
|
|
|
|
|
|
|
|
|
|
|
|
|
| 544 |
except Exception as e:
|
| 545 |
print(f"์ปจํ
์คํธ ๊ฒ์ ์ค๋ฅ: {str(e)}")
|
| 546 |
wiki_context = ""
|
| 547 |
-
|
| 548 |
# ๋ํ ํ์คํ ๋ฆฌ ๊ตฌ์ฑ
|
| 549 |
conversation = []
|
| 550 |
for prompt, answer in history:
|
|
@@ -557,36 +652,63 @@ def stream_chat(message: str, history: list, uploaded_file, temperature: float,
|
|
| 557 |
final_message = file_context + wiki_context + "\nํ์ฌ ์ง๋ฌธ: " + message
|
| 558 |
conversation.append({"role": "user", "content": final_message})
|
| 559 |
|
| 560 |
-
#
|
| 561 |
-
|
| 562 |
-
|
| 563 |
-
|
| 564 |
-
|
| 565 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 566 |
|
| 567 |
-
inputs = tokenizer(input_ids, return_tensors="pt").to("cuda")
|
| 568 |
-
|
| 569 |
-
# ๋ฉ๋ชจ๋ฆฌ ์ฌ์ฉ๋ ์ฒดํฌ
|
| 570 |
if torch.cuda.is_available():
|
| 571 |
print(f"์
๋ ฅ ํ
์ ์์ฑ ํ CUDA ๋ฉ๋ชจ๋ฆฌ: {torch.cuda.memory_allocated() / 1024**2:.2f} MB")
|
| 572 |
|
| 573 |
-
streamer = TextIteratorStreamer(
|
|
|
|
|
|
|
| 574 |
|
| 575 |
generate_kwargs = dict(
|
| 576 |
-
inputs,
|
| 577 |
streamer=streamer,
|
| 578 |
top_k=top_k,
|
| 579 |
top_p=top_p,
|
| 580 |
repetition_penalty=penalty,
|
| 581 |
-
max_new_tokens=
|
| 582 |
-
do_sample=True,
|
| 583 |
temperature=temperature,
|
| 584 |
-
eos_token_id=
|
| 585 |
)
|
| 586 |
-
|
| 587 |
# ์์ฑ ์์ ์ ๋ฉ๋ชจ๋ฆฌ ์ ๋ฆฌ
|
| 588 |
clear_cuda_memory()
|
| 589 |
-
|
| 590 |
thread = Thread(target=model.generate, kwargs=generate_kwargs)
|
| 591 |
thread.start()
|
| 592 |
|
|
@@ -601,12 +723,10 @@ def stream_chat(message: str, history: list, uploaded_file, temperature: float,
|
|
| 601 |
except Exception as e:
|
| 602 |
error_message = f"์ค๋ฅ๊ฐ ๋ฐ์ํ์ต๋๋ค: {str(e)}"
|
| 603 |
print(f"Stream chat ์ค๋ฅ: {error_message}")
|
| 604 |
-
# ์ค๋ฅ ๋ฐ์ ์์๋ ๋ฉ๋ชจ๋ฆฌ ์ ๋ฆฌ
|
| 605 |
clear_cuda_memory()
|
| 606 |
yield "", history + [[message, error_message]]
|
| 607 |
|
| 608 |
|
| 609 |
-
|
| 610 |
def create_demo():
|
| 611 |
with gr.Blocks(css=CSS) as demo:
|
| 612 |
with gr.Column(elem_classes="markdown-style"):
|
|
@@ -615,14 +735,14 @@ def create_demo():
|
|
| 615 |
#### ๐ RAG: Upload and Analyze Files (TXT, CSV, PDF, Parquet files)
|
| 616 |
Upload your files for data analysis and learning
|
| 617 |
""")
|
| 618 |
-
|
| 619 |
chatbot = gr.Chatbot(
|
| 620 |
value=[],
|
| 621 |
height=600,
|
| 622 |
label="GiniGEN AI Assistant",
|
| 623 |
elem_classes="chat-container"
|
| 624 |
)
|
| 625 |
-
|
| 626 |
with gr.Row(elem_classes="input-container"):
|
| 627 |
with gr.Column(scale=1, min_width=70):
|
| 628 |
file_upload = gr.File(
|
|
@@ -633,7 +753,7 @@ def create_demo():
|
|
| 633 |
interactive=True,
|
| 634 |
show_label=False
|
| 635 |
)
|
| 636 |
-
|
| 637 |
with gr.Column(scale=3):
|
| 638 |
msg = gr.Textbox(
|
| 639 |
show_label=False,
|
|
@@ -642,21 +762,21 @@ def create_demo():
|
|
| 642 |
elem_classes="input-textbox",
|
| 643 |
scale=1
|
| 644 |
)
|
| 645 |
-
|
| 646 |
with gr.Column(scale=1, min_width=70):
|
| 647 |
send = gr.Button(
|
| 648 |
"Send",
|
| 649 |
elem_classes="send-button custom-button",
|
| 650 |
scale=1
|
| 651 |
)
|
| 652 |
-
|
| 653 |
with gr.Column(scale=1, min_width=70):
|
| 654 |
clear = gr.Button(
|
| 655 |
"Clear",
|
| 656 |
elem_classes="clear-button custom-button",
|
| 657 |
scale=1
|
| 658 |
)
|
| 659 |
-
|
| 660 |
with gr.Accordion("๐ฎ Advanced Settings", open=False):
|
| 661 |
with gr.Row():
|
| 662 |
with gr.Column(scale=1):
|
|
@@ -697,7 +817,7 @@ def create_demo():
|
|
| 697 |
current_file_context = None
|
| 698 |
return [], None, "Start a new conversation..."
|
| 699 |
|
| 700 |
-
#
|
| 701 |
msg.submit(
|
| 702 |
stream_chat,
|
| 703 |
inputs=[msg, chatbot, file_upload, temperature, max_new_tokens, top_p, top_k, penalty],
|
|
@@ -721,7 +841,6 @@ def create_demo():
|
|
| 721 |
queue=True
|
| 722 |
)
|
| 723 |
|
| 724 |
-
# Clear button event binding
|
| 725 |
clear.click(
|
| 726 |
fn=clear_conversation,
|
| 727 |
outputs=[chatbot, file_upload, msg],
|
|
@@ -730,6 +849,7 @@ def create_demo():
|
|
| 730 |
|
| 731 |
return demo
|
| 732 |
|
|
|
|
| 733 |
if __name__ == "__main__":
|
| 734 |
demo = create_demo()
|
| 735 |
-
demo.launch()
|
|
|
|
| 1 |
+
import os
|
| 2 |
+
# Dynamo ์์ ๋นํ์ฑํ
|
| 3 |
+
os.environ["TORCH_DYNAMO_DISABLE"] = "1"
|
| 4 |
+
|
| 5 |
import torch
|
| 6 |
+
import torch._dynamo
|
| 7 |
import gradio as gr
|
| 8 |
import spaces
|
| 9 |
from transformers import AutoModelForCausalLM, AutoTokenizer, TextIteratorStreamer
|
| 10 |
|
|
|
|
| 11 |
from threading import Thread
|
| 12 |
import random
|
| 13 |
from datasets import load_dataset
|
|
|
|
| 21 |
import pypdf
|
| 22 |
import io
|
| 23 |
import pyarrow.parquet as pq
|
| 24 |
+
from tabulate import tabulate
|
|
|
|
|
|
|
| 25 |
import platform
|
| 26 |
import subprocess
|
| 27 |
import pytesseract
|
| 28 |
from pdf2image import convert_from_path
|
| 29 |
|
| 30 |
+
# -------------------- ์ถ๊ฐ: PDF to Markdown ๋ณํ ๊ด๋ จ import --------------------
|
| 31 |
+
import re
|
| 32 |
+
import requests
|
| 33 |
+
from bs4 import BeautifulSoup
|
| 34 |
+
import urllib.request
|
| 35 |
+
import ocrmypdf
|
| 36 |
+
import pytz
|
| 37 |
+
import urllib.parse
|
| 38 |
+
from pypdf import PdfReader
|
| 39 |
+
# ---------------------------------------------------------------------------
|
| 40 |
+
|
| 41 |
+
# --------------------
|
| 42 |
+
# 1) Dynamo suppress_errors ์ต์
์ฌ์ฉ (์ค๋ฅ ์ eager๋ก fallback)
|
| 43 |
+
# --------------------
|
| 44 |
+
torch._dynamo.config.suppress_errors = True
|
| 45 |
+
|
| 46 |
+
# ์ ์ญ ๋ณ์
|
| 47 |
current_file_context = None
|
| 48 |
|
| 49 |
# ํ๊ฒฝ ๋ณ์ ์ค์
|
|
|
|
| 66 |
question_vectors = vectorizer.fit_transform(questions)
|
| 67 |
print("TF-IDF ๋ฒกํฐํ ์๋ฃ")
|
| 68 |
|
| 69 |
+
|
| 70 |
class ChatHistory:
|
| 71 |
def __init__(self):
|
| 72 |
self.history = []
|
|
|
|
| 122 |
print(f"ํ์คํ ๋ฆฌ ๋ก๋ ์คํจ: {e}")
|
| 123 |
self.history = []
|
| 124 |
|
| 125 |
+
|
| 126 |
# ์ ์ญ ChatHistory ์ธ์คํด์ค ์์ฑ
|
| 127 |
chat_history = ChatHistory()
|
| 128 |
|
| 129 |
+
|
| 130 |
def find_relevant_context(query, top_k=3):
|
| 131 |
# ์ฟผ๋ฆฌ ๋ฒกํฐํ
|
| 132 |
query_vector = vectorizer.transform([query])
|
|
|
|
| 133 |
# ์ฝ์ฌ์ธ ์ ์ฌ๋ ๊ณ์ฐ
|
| 134 |
similarities = (query_vector * question_vectors.T).toarray()[0]
|
|
|
|
| 135 |
# ๊ฐ์ฅ ์ ์ฌํ ์ง๋ฌธ๋ค์ ์ธ๋ฑ์ค
|
| 136 |
top_indices = np.argsort(similarities)[-top_k:][::-1]
|
|
|
|
| 137 |
# ๊ด๋ จ ์ปจํ
์คํธ ์ถ์ถ
|
| 138 |
relevant_contexts = []
|
| 139 |
for idx in top_indices:
|
|
|
|
| 143 |
'answer': wiki_dataset['train']['answer'][idx],
|
| 144 |
'similarity': similarities[idx]
|
| 145 |
})
|
|
|
|
| 146 |
return relevant_contexts
|
| 147 |
|
| 148 |
+
|
| 149 |
def init_msg():
|
| 150 |
+
return "ํ์ผ์ ๋ถ์ํ๊ณ ์์ต๋๋ค..."
|
| 151 |
+
|
| 152 |
+
|
| 153 |
+
# -------------------- PDF ํ์ผ์ Markdown์ผ๋ก ๋ณํํ๋ ์ ํธ ํจ์๋ค --------------------
|
| 154 |
+
def extract_text_from_pdf(reader: PdfReader) -> str:
|
| 155 |
+
"""
|
| 156 |
+
PyPDF๋ฅผ ์ฌ์ฉํด ๋ชจ๋ ํ์ด์ง ํ
์คํธ๋ฅผ ์ถ์ถ.
|
| 157 |
+
๋ง์ฝ ํ
์คํธ๊ฐ ์์ผ๋ฉด ๋น ๋ฌธ์์ด ๋ฐํ.
|
| 158 |
+
"""
|
| 159 |
+
full_text = ""
|
| 160 |
+
for idx, page in enumerate(reader.pages):
|
| 161 |
+
text = page.extract_text() or ""
|
| 162 |
+
if len(text) > 0:
|
| 163 |
+
full_text += f"---- Page {idx+1} ----\n" + text + "\n\n"
|
| 164 |
+
return full_text.strip()
|
| 165 |
+
|
| 166 |
+
|
| 167 |
+
def convert_pdf_to_markdown(pdf_file: str):
|
| 168 |
+
"""
|
| 169 |
+
PDF ํ์ผ์ ์ฝ๊ณ ํ
์คํธ๋ฅผ ์ถ์ถํ ๋ค,
|
| 170 |
+
์ด๋ฏธ์ง๊ฐ ๋ง๊ณ ํ
์คํธ๊ฐ ์ ์ ๊ฒฝ์ฐ์๋ OCR์ ์๋ํ๋ค.
|
| 171 |
+
์ต์ข
์ ์ผ๋ก Markdown ํ์์ผ๋ก ๋ณํ ๊ฐ๋ฅํ ํ
์คํธ๋ฅผ ๋ฐํํ๋ค.
|
| 172 |
+
๋ฉํ๋ฐ์ดํฐ๋ ํจ๊ป ๋ฐํ.
|
| 173 |
+
"""
|
| 174 |
+
try:
|
| 175 |
+
reader = PdfReader(pdf_file)
|
| 176 |
+
except Exception as e:
|
| 177 |
+
return f"PDF ํ์ผ์ ์ฝ๋ ์ค ์ค๋ฅ ๋ฐ์: {e}", None, None
|
| 178 |
+
|
| 179 |
+
# Extract metadata
|
| 180 |
+
raw_meta = reader.metadata
|
| 181 |
+
metadata = {
|
| 182 |
+
"author": raw_meta.author if raw_meta else None,
|
| 183 |
+
"creator": raw_meta.creator if raw_meta else None,
|
| 184 |
+
"producer": raw_meta.producer if raw_meta else None,
|
| 185 |
+
"subject": raw_meta.subject if raw_meta else None,
|
| 186 |
+
"title": raw_meta.title if raw_meta else None,
|
| 187 |
+
}
|
| 188 |
+
|
| 189 |
+
# Extract text
|
| 190 |
+
full_text = extract_text_from_pdf(reader)
|
| 191 |
+
|
| 192 |
+
# ์ด๋ฏธ์ง๊ฐ ๋ง๊ณ ํ
์คํธ๊ฐ ๋๋ฌด ์งง์ผ๋ฉด OCR ์๋
|
| 193 |
+
image_count = 0
|
| 194 |
+
for page in reader.pages:
|
| 195 |
+
image_count += len(page.images)
|
| 196 |
+
|
| 197 |
+
if image_count > 0 and len(full_text) < 1000:
|
| 198 |
+
try:
|
| 199 |
+
out_pdf_file = pdf_file.replace(".pdf", "_ocr.pdf")
|
| 200 |
+
ocrmypdf.ocr(pdf_file, out_pdf_file, force_ocr=True)
|
| 201 |
+
# Re-extract text from OCR-processed PDF
|
| 202 |
+
reader_ocr = PdfReader(out_pdf_file)
|
| 203 |
+
full_text = extract_text_from_pdf(reader_ocr)
|
| 204 |
+
except Exception as e:
|
| 205 |
+
full_text = f"OCR ์ฒ๋ฆฌ ์ค ์ค๋ฅ ๋ฐ์: {e}\n\n์๋ณธ PDF ํ
์คํธ:\n\n" + full_text
|
| 206 |
+
|
| 207 |
+
return full_text, metadata, pdf_file
|
| 208 |
+
|
| 209 |
+
|
| 210 |
+
# ---------------------------------------------------------------------------
|
| 211 |
|
| 212 |
def analyze_file_content(content, file_type):
|
| 213 |
+
"""ํ์ผ ๋ด์ฉ์ ๊ฐ๋จํ ๋ถ์ํ ํ ๊ตฌ์กฐ ์์ฝ์ ๋ฐํ."""
|
| 214 |
if file_type in ['parquet', 'csv']:
|
| 215 |
try:
|
| 216 |
lines = content.split('\n')
|
|
|
|
| 220 |
return f"๐ Dataset Structure: {columns} columns, {rows} rows"
|
| 221 |
except:
|
| 222 |
return "โ Failed to analyze dataset structure"
|
| 223 |
+
|
| 224 |
lines = content.split('\n')
|
| 225 |
total_lines = len(lines)
|
| 226 |
non_empty_lines = len([line for line in lines if line.strip()])
|
| 227 |
+
|
| 228 |
if any(keyword in content.lower() for keyword in ['def ', 'class ', 'import ', 'function']):
|
| 229 |
functions = len([line for line in lines if 'def ' in line])
|
| 230 |
classes = len([line for line in lines if 'class ' in line])
|
| 231 |
imports = len([line for line in lines if 'import ' in line or 'from ' in line])
|
| 232 |
return f"๐ป Code Structure: {total_lines} lines (Functions: {functions}, Classes: {classes}, Imports: {imports})"
|
| 233 |
+
|
| 234 |
paragraphs = content.count('\n\n') + 1
|
| 235 |
words = len(content.split())
|
| 236 |
return f"๐ Document Structure: {total_lines} lines, {paragraphs} paragraphs, approximately {words} words"
|
| 237 |
|
| 238 |
+
|
| 239 |
def read_uploaded_file(file):
|
| 240 |
+
"""
|
| 241 |
+
์
๋ก๋๋ ํ์ผ์ ์ฒ๋ฆฌํ์ฌ
|
| 242 |
+
1) ํ์ผ ํ์
๋ณ๋ก ๋ด์ฉ์ ์ฝ๊ณ
|
| 243 |
+
2) ๋ถ์ ๊ฒฐ๊ณผ์ ํจ๊ป ๋ฐํ
|
| 244 |
+
"""
|
| 245 |
if file is None:
|
| 246 |
return "", ""
|
| 247 |
try:
|
| 248 |
file_ext = os.path.splitext(file.name)[1].lower()
|
| 249 |
+
|
| 250 |
+
# Parquet
|
| 251 |
if file_ext == '.parquet':
|
| 252 |
try:
|
| 253 |
table = pq.read_table(file.name)
|
| 254 |
df = table.to_pandas()
|
| 255 |
+
|
| 256 |
content = f"๐ Parquet File Analysis:\n\n"
|
| 257 |
content += f"1. Basic Information:\n"
|
| 258 |
content += f"- Total Rows: {len(df):,}\n"
|
| 259 |
content += f"- Total Columns: {len(df.columns)}\n"
|
| 260 |
content += f"- Memory Usage: {df.memory_usage(deep=True).sum() / 1024 / 1024:.2f} MB\n\n"
|
| 261 |
+
|
| 262 |
content += f"2. Column Information:\n"
|
| 263 |
for col in df.columns:
|
| 264 |
content += f"- {col} ({df[col].dtype})\n"
|
| 265 |
+
|
| 266 |
content += f"\n3. Data Preview:\n"
|
| 267 |
content += tabulate(df.head(5), headers='keys', tablefmt='pipe', showindex=False)
|
| 268 |
+
|
| 269 |
content += f"\n\n4. Missing Values:\n"
|
| 270 |
null_counts = df.isnull().sum()
|
| 271 |
for col, count in null_counts[null_counts > 0].items():
|
| 272 |
content += f"- {col}: {count:,} ({count/len(df)*100:.1f}%)\n"
|
| 273 |
+
|
| 274 |
numeric_cols = df.select_dtypes(include=['int64', 'float64']).columns
|
| 275 |
if len(numeric_cols) > 0:
|
| 276 |
content += f"\n5. Numeric Column Statistics:\n"
|
| 277 |
stats_df = df[numeric_cols].describe()
|
| 278 |
content += tabulate(stats_df, headers='keys', tablefmt='pipe')
|
| 279 |
+
|
| 280 |
return content, "parquet"
|
| 281 |
except Exception as e:
|
| 282 |
return f"Error reading Parquet file: {str(e)}", "error"
|
| 283 |
+
|
| 284 |
+
# PDF (Markdown ๋ณํ)
|
| 285 |
if file_ext == '.pdf':
|
| 286 |
try:
|
| 287 |
+
markdown_text, metadata, processed_pdf_path = convert_pdf_to_markdown(file.name)
|
| 288 |
+
if metadata is None:
|
| 289 |
+
return f"PDF ํ์ผ ๋ณํ ์ค๋ฅ ๋๋ ์ฝ๊ธฐ ์คํจ.\n\n์๋ณธ ๋ฉ์์ง:\n{markdown_text}", "error"
|
| 290 |
+
|
| 291 |
+
content = "# PDF to Markdown Conversion\n\n"
|
| 292 |
+
content += "## Metadata\n"
|
| 293 |
+
for k, v in metadata.items():
|
| 294 |
+
content += f"**{k.capitalize()}**: {v}\n\n"
|
| 295 |
+
|
| 296 |
+
content += "## Extracted Text\n\n"
|
| 297 |
+
content += markdown_text
|
| 298 |
+
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 299 |
return content, "pdf"
|
| 300 |
except Exception as e:
|
| 301 |
return f"Error reading PDF file: {str(e)}", "error"
|
| 302 |
+
|
| 303 |
+
# CSV
|
| 304 |
elif file_ext == '.csv':
|
| 305 |
encodings = ['utf-8', 'cp949', 'euc-kr', 'latin1']
|
| 306 |
for encoding in encodings:
|
|
|
|
| 311 |
content += f"- Total Rows: {len(df):,}\n"
|
| 312 |
content += f"- Total Columns: {len(df.columns)}\n"
|
| 313 |
content += f"- Memory Usage: {df.memory_usage(deep=True).sum() / 1024 / 1024:.2f} MB\n\n"
|
| 314 |
+
|
| 315 |
content += f"2. Column Information:\n"
|
| 316 |
for col in df.columns:
|
| 317 |
content += f"- {col} ({df[col].dtype})\n"
|
| 318 |
+
|
| 319 |
content += f"\n3. Data Preview:\n"
|
| 320 |
content += df.head(5).to_markdown(index=False)
|
| 321 |
+
|
| 322 |
content += f"\n\n4. Missing Values:\n"
|
| 323 |
null_counts = df.isnull().sum()
|
| 324 |
for col, count in null_counts[null_counts > 0].items():
|
| 325 |
content += f"- {col}: {count:,} ({count/len(df)*100:.1f}%)\n"
|
| 326 |
+
|
| 327 |
return content, "csv"
|
| 328 |
except UnicodeDecodeError:
|
| 329 |
continue
|
| 330 |
raise UnicodeDecodeError(f"Unable to read file with supported encodings ({', '.join(encodings)})")
|
| 331 |
+
|
| 332 |
+
# ์ผ๋ฐ ํ
์คํธ ํ์ผ
|
| 333 |
else:
|
| 334 |
encodings = ['utf-8', 'cp949', 'euc-kr', 'latin1']
|
| 335 |
for encoding in encodings:
|
| 336 |
try:
|
| 337 |
with open(file.name, 'r', encoding=encoding) as f:
|
| 338 |
content = f.read()
|
| 339 |
+
|
| 340 |
lines = content.split('\n')
|
| 341 |
total_lines = len(lines)
|
| 342 |
non_empty_lines = len([line for line in lines if line.strip()])
|
| 343 |
+
|
| 344 |
is_code = any(keyword in content.lower() for keyword in ['def ', 'class ', 'import ', 'function'])
|
| 345 |
+
|
| 346 |
analysis = f"\n๐ File Analysis:\n"
|
| 347 |
if is_code:
|
| 348 |
functions = len([line for line in lines if 'def ' in line])
|
| 349 |
classes = len([line for line in lines if 'class ' in line])
|
| 350 |
imports = len([line for line in lines if 'import ' in line or 'from ' in line])
|
| 351 |
+
|
| 352 |
analysis += f"- File Type: Code\n"
|
| 353 |
analysis += f"- Total Lines: {total_lines:,}\n"
|
| 354 |
analysis += f"- Functions: {functions}\n"
|
|
|
|
| 357 |
else:
|
| 358 |
words = len(content.split())
|
| 359 |
chars = len(content)
|
| 360 |
+
|
| 361 |
analysis += f"- File Type: Text\n"
|
| 362 |
analysis += f"- Total Lines: {total_lines:,}\n"
|
| 363 |
analysis += f"- Non-empty Lines: {non_empty_lines:,}\n"
|
| 364 |
analysis += f"- Word Count: {words:,}\n"
|
| 365 |
analysis += f"- Character Count: {chars:,}\n"
|
| 366 |
+
|
| 367 |
return content + analysis, "text"
|
| 368 |
except UnicodeDecodeError:
|
| 369 |
continue
|
| 370 |
raise UnicodeDecodeError(f"Unable to read file with supported encodings ({', '.join(encodings)})")
|
| 371 |
+
|
| 372 |
except Exception as e:
|
| 373 |
return f"Error reading file: {str(e)}", "error"
|
| 374 |
|
|
|
|
| 511 |
font-size: 1.1em !important;
|
| 512 |
padding: 10px 15px !important;
|
| 513 |
display: flex !important;
|
| 514 |
+
align-items: flex-start !important;
|
| 515 |
}
|
| 516 |
.input-textbox textarea {
|
| 517 |
+
padding-top: 5px !important;
|
| 518 |
}
|
| 519 |
.send-button {
|
| 520 |
height: 70px !important;
|
|
|
|
| 528 |
}
|
| 529 |
"""
|
| 530 |
|
|
|
|
| 531 |
def clear_cuda_memory():
|
| 532 |
if hasattr(torch.cuda, 'empty_cache'):
|
| 533 |
with torch.cuda.device('cuda'):
|
| 534 |
torch.cuda.empty_cache()
|
| 535 |
|
| 536 |
+
|
| 537 |
@spaces.GPU
|
| 538 |
def load_model():
|
| 539 |
try:
|
| 540 |
+
loaded_model = AutoModelForCausalLM.from_pretrained(
|
| 541 |
MODEL_ID,
|
| 542 |
torch_dtype=torch.bfloat16,
|
| 543 |
device_map="auto",
|
| 544 |
)
|
| 545 |
+
return loaded_model
|
| 546 |
except Exception as e:
|
| 547 |
print(f"๋ชจ๋ธ ๋ก๋ ์ค๋ฅ: {str(e)}")
|
| 548 |
raise
|
| 549 |
|
| 550 |
+
def _truncate_tokens_for_context(input_ids_str: str, desired_input_length: int) -> str:
|
| 551 |
+
"""
|
| 552 |
+
์
๋ ฅ ๋ฌธ์์ด์ด desired_input_length ํ ํฐ์ ๋์ผ๋ฉด, ์๋ถ๋ถ(์ค๋๋ ์ปจํ
์คํธ)์ ์๋ผ๋ด๋ ํจ์.
|
| 553 |
+
"""
|
| 554 |
+
tokens = input_ids_str.split()
|
| 555 |
+
if len(tokens) > desired_input_length:
|
| 556 |
+
# ๊ฐ์ฅ ์ค๋๋ ๋ถ๋ถ์ ๋ฒ๋ฆฌ๊ณ , ๋ค์์ desired_input_length๋ง ๋จ๊น
|
| 557 |
+
tokens = tokens[-desired_input_length:]
|
| 558 |
+
return " ".join(tokens)
|
| 559 |
+
|
| 560 |
+
|
| 561 |
+
# build_prompt ํจ์: ๋ํ ๋ด์ญ์ ๋ฌธ์์ด๋ก ๋ณํ
|
| 562 |
+
def build_prompt(conversation: list) -> str:
|
| 563 |
+
"""
|
| 564 |
+
conversation์ ๊ฐ ํญ๋ชฉ์ด {"role": "user" ๋๋ "assistant", "content": ...} ํํ์ ๋์
๋๋ฆฌ ๋ชฉ๋ก์
๋๋ค.
|
| 565 |
+
์ด๋ฅผ ๋จ์ ํ
์คํธ ํ๋กฌํํธ๋ก ๋ณํํฉ๋๋ค.
|
| 566 |
+
"""
|
| 567 |
+
prompt = ""
|
| 568 |
+
for msg in conversation:
|
| 569 |
+
if msg["role"] == "user":
|
| 570 |
+
prompt += "User: " + msg["content"] + "\n"
|
| 571 |
+
elif msg["role"] == "assistant":
|
| 572 |
+
prompt += "Assistant: " + msg["content"] + "\n"
|
| 573 |
+
# ๋ง์ง๋ง์ ์ด์์คํดํธ ์๋ต์ ๊ธฐ๋ํ๋๋ก ์ถ๊ฐ
|
| 574 |
+
prompt += "Assistant: "
|
| 575 |
+
return prompt
|
| 576 |
+
|
| 577 |
+
|
| 578 |
@spaces.GPU
|
| 579 |
+
def stream_chat(
|
| 580 |
+
message: str,
|
| 581 |
+
history: list,
|
| 582 |
+
uploaded_file,
|
| 583 |
+
temperature: float,
|
| 584 |
+
max_new_tokens: int,
|
| 585 |
+
top_p: float,
|
| 586 |
+
top_k: int,
|
| 587 |
+
penalty: float
|
| 588 |
+
):
|
| 589 |
global model, current_file_context
|
| 590 |
+
|
| 591 |
try:
|
| 592 |
if model is None:
|
| 593 |
model = load_model()
|
| 594 |
+
|
| 595 |
print(f'message is - {message}')
|
| 596 |
print(f'history is - {history}')
|
| 597 |
|
| 598 |
# ํ์ผ ์
๋ก๋ ์ฒ๋ฆฌ
|
| 599 |
file_context = ""
|
| 600 |
if uploaded_file and message == "ํ์ผ์ ๋ถ์ํ๊ณ ์์ต๋๋ค...":
|
| 601 |
+
# ์๋ก์ด ํ์ผ ์
๋ก๋ ์์๋ ๊ธฐ์กด ๋ฉ๋ชจ๋ฆฌ ์ปจํ
์คํธ ์ด๊ธฐํ
|
| 602 |
+
current_file_context = None
|
| 603 |
try:
|
| 604 |
content, file_type = read_uploaded_file(uploaded_file)
|
| 605 |
if content:
|
| 606 |
file_analysis = analyze_file_content(content, file_type)
|
| 607 |
+
file_context = (
|
| 608 |
+
f"\n\n๐ ํ์ผ ๋ถ์ ๊ฒฐ๊ณผ:\n{file_analysis}"
|
| 609 |
+
f"\n\nํ์ผ ๋ด์ฉ:\n```\n{content}\n```"
|
| 610 |
+
)
|
| 611 |
current_file_context = file_context # ํ์ผ ์ปจํ
์คํธ ์ ์ฅ
|
| 612 |
message = "์
๋ก๋๋ ํ์ผ์ ๋ถ์ํด์ฃผ์ธ์."
|
| 613 |
except Exception as e:
|
| 614 |
print(f"ํ์ผ ๋ถ์ ์ค๋ฅ: {str(e)}")
|
| 615 |
file_context = f"\n\nโ ํ์ผ ๋ถ์ ์ค ์ค๋ฅ๊ฐ ๋ฐ์ํ์ต๋๋ค: {str(e)}"
|
| 616 |
+
elif current_file_context:
|
| 617 |
+
# ์ด๋ฏธ ์
๋ก๋๋ ํ์ผ ์ปจํ
์คํธ๊ฐ ์๋ค๋ฉด ์ฌ์ฉ
|
| 618 |
file_context = current_file_context
|
|
|
|
| 619 |
|
| 620 |
# ๋ฉ๋ชจ๋ฆฌ ์ฌ์ฉ๋ ๋ชจ๋ํฐ๋ง
|
| 621 |
if torch.cuda.is_available():
|
| 622 |
print(f"CUDA ๋ฉ๋ชจ๋ฆฌ ์ฌ์ฉ๋: {torch.cuda.memory_allocated() / 1024**2:.2f} MB")
|
| 623 |
|
| 624 |
# ๋ํ ํ์คํ ๋ฆฌ๊ฐ ๋๋ฌด ๊ธธ๋ฉด ์๋ผ๋ด๊ธฐ
|
| 625 |
+
max_history_length = 10
|
| 626 |
if len(history) > max_history_length:
|
| 627 |
history = history[-max_history_length:]
|
| 628 |
|
| 629 |
+
# ์ํค ์ปจํ
์คํธ ์ฐพ๊ธฐ
|
| 630 |
try:
|
| 631 |
relevant_contexts = find_relevant_context(message)
|
| 632 |
wiki_context = "\n\n๊ด๋ จ ์ํคํผ๋์ ์ ๋ณด:\n"
|
| 633 |
for ctx in relevant_contexts:
|
| 634 |
+
wiki_context += (
|
| 635 |
+
f"Q: {ctx['question']}\n"
|
| 636 |
+
f"A: {ctx['answer']}\n"
|
| 637 |
+
f"์ ์ฌ๋: {ctx['similarity']:.3f}\n\n"
|
| 638 |
+
)
|
| 639 |
except Exception as e:
|
| 640 |
print(f"์ปจํ
์คํธ ๊ฒ์ ์ค๋ฅ: {str(e)}")
|
| 641 |
wiki_context = ""
|
| 642 |
+
|
| 643 |
# ๋ํ ํ์คํ ๋ฆฌ ๊ตฌ์ฑ
|
| 644 |
conversation = []
|
| 645 |
for prompt, answer in history:
|
|
|
|
| 652 |
final_message = file_context + wiki_context + "\nํ์ฌ ์ง๋ฌธ: " + message
|
| 653 |
conversation.append({"role": "user", "content": final_message})
|
| 654 |
|
| 655 |
+
# build_prompt ์ฌ์ฉ (๊ธฐ์กด tokenizer.apply_chat_template ๋์ )
|
| 656 |
+
input_ids_str = build_prompt(conversation)
|
| 657 |
+
# ๋จผ์ 6000 ํ ํฐ ์ด๋ด๋ก ์๋ผ์ฃผ๊ธฐ (์์์ ์์น, ํ์์ ๋ฐ๋ผ ์กฐ์ ๊ฐ๋ฅ)
|
| 658 |
+
input_ids_str = _truncate_tokens_for_context(input_ids_str, 6000)
|
| 659 |
+
|
| 660 |
+
inputs = tokenizer(input_ids_str, return_tensors="pt").to("cuda")
|
| 661 |
+
|
| 662 |
+
# ์ต๋ ์ปจํ
์คํธ 8192 ๊ณ ๋ คํ์ฌ, ๋จ์ ์๋ฆฌ๊ฐ ์ ์ผ๋ฉด max_new_tokens ์ค์ด๊ธฐ
|
| 663 |
+
max_context = 8192
|
| 664 |
+
input_length = inputs["input_ids"].shape[1]
|
| 665 |
+
remaining = max_context - input_length
|
| 666 |
+
|
| 667 |
+
# ์ต์ 128 ํ ํฐ ์ ๋๋ ์์ฑํ ์ ์๊ฒ ๋ง๋ค๊ณ ์ถ๋ค๋ฉด,
|
| 668 |
+
# remaining์ด 128 ๋ฏธ๋ง์ด๋ฉด, ์ถ๊ฐ๋ก input์ ๋ ์๋ผ๋ธ๋ค.
|
| 669 |
+
min_generation = 128
|
| 670 |
+
if remaining < min_generation:
|
| 671 |
+
# ๋ ์๋ผ์ ์ถฉ๋ถํ ์ถ๋ ฅ ํ ํฐ ํ๋ณด
|
| 672 |
+
must_cut = min_generation - remaining # ๋ช ํ ํฐ๋งํผ ๋ ์๋ฅผ์ง
|
| 673 |
+
new_desired_input_length = max(1, input_length - must_cut)
|
| 674 |
+
print(f"[์ฃผ์] ์
๋ ฅ์ด ๋๋ฌด ๊ธธ์ด {must_cut}ํ ํฐ ๋ ์ ๊ฑฐํ์ฌ, input_length={input_length} -> {new_desired_input_length} ์ฌ์กฐ์ ")
|
| 675 |
+
# ๋ฌธ์์ด ๋ค์ ๋ง๋ค์ด์ tokenizer
|
| 676 |
+
input_ids_str = _truncate_tokens_for_context(input_ids_str, new_desired_input_length)
|
| 677 |
+
inputs = tokenizer(input_ids_str, return_tensors="pt").to("cuda")
|
| 678 |
+
input_length = inputs["input_ids"].shape[1]
|
| 679 |
+
remaining = max_context - input_length
|
| 680 |
+
|
| 681 |
+
# ์ต์ข
์ ์ผ๋ก (input + max_new_tokens) <= 8192 ๋๋๋ก
|
| 682 |
+
if remaining < max_new_tokens:
|
| 683 |
+
print(f"[์ฃผ์] ์
๋ ฅ ํ ํฐ์ด ๋ง์ max_new_tokens={max_new_tokens} -> {remaining}๋ก ์กฐ์ ํฉ๋๋ค.")
|
| 684 |
+
max_new_tokens = remaining
|
| 685 |
+
|
| 686 |
+
if max_new_tokens < 1:
|
| 687 |
+
# ๊ทธ๋๋ 1 ๋ฏธ๋ง์ด๋ฉด 1 ํ ํฐ๋ง ์์ฑ
|
| 688 |
+
max_new_tokens = 1
|
| 689 |
|
|
|
|
|
|
|
|
|
|
| 690 |
if torch.cuda.is_available():
|
| 691 |
print(f"์
๋ ฅ ํ
์ ์์ฑ ํ CUDA ๋ฉ๋ชจ๋ฆฌ: {torch.cuda.memory_allocated() / 1024**2:.2f} MB")
|
| 692 |
|
| 693 |
+
streamer = TextIteratorStreamer(
|
| 694 |
+
tokenizer, timeout=10., skip_prompt=True, skip_special_tokens=True
|
| 695 |
+
)
|
| 696 |
|
| 697 |
generate_kwargs = dict(
|
| 698 |
+
**inputs,
|
| 699 |
streamer=streamer,
|
| 700 |
top_k=top_k,
|
| 701 |
top_p=top_p,
|
| 702 |
repetition_penalty=penalty,
|
| 703 |
+
max_new_tokens=max_new_tokens,
|
| 704 |
+
do_sample=True,
|
| 705 |
temperature=temperature,
|
| 706 |
+
eos_token_id=255001, # ์์ : ๋ฆฌ์คํธ ๋์ ์ ์ํ ์ฌ์ฉ
|
| 707 |
)
|
| 708 |
+
|
| 709 |
# ์์ฑ ์์ ์ ๋ฉ๋ชจ๋ฆฌ ์ ๋ฆฌ
|
| 710 |
clear_cuda_memory()
|
| 711 |
+
|
| 712 |
thread = Thread(target=model.generate, kwargs=generate_kwargs)
|
| 713 |
thread.start()
|
| 714 |
|
|
|
|
| 723 |
except Exception as e:
|
| 724 |
error_message = f"์ค๋ฅ๊ฐ ๋ฐ์ํ์ต๋๋ค: {str(e)}"
|
| 725 |
print(f"Stream chat ์ค๋ฅ: {error_message}")
|
|
|
|
| 726 |
clear_cuda_memory()
|
| 727 |
yield "", history + [[message, error_message]]
|
| 728 |
|
| 729 |
|
|
|
|
| 730 |
def create_demo():
|
| 731 |
with gr.Blocks(css=CSS) as demo:
|
| 732 |
with gr.Column(elem_classes="markdown-style"):
|
|
|
|
| 735 |
#### ๐ RAG: Upload and Analyze Files (TXT, CSV, PDF, Parquet files)
|
| 736 |
Upload your files for data analysis and learning
|
| 737 |
""")
|
| 738 |
+
|
| 739 |
chatbot = gr.Chatbot(
|
| 740 |
value=[],
|
| 741 |
height=600,
|
| 742 |
label="GiniGEN AI Assistant",
|
| 743 |
elem_classes="chat-container"
|
| 744 |
)
|
| 745 |
+
|
| 746 |
with gr.Row(elem_classes="input-container"):
|
| 747 |
with gr.Column(scale=1, min_width=70):
|
| 748 |
file_upload = gr.File(
|
|
|
|
| 753 |
interactive=True,
|
| 754 |
show_label=False
|
| 755 |
)
|
| 756 |
+
|
| 757 |
with gr.Column(scale=3):
|
| 758 |
msg = gr.Textbox(
|
| 759 |
show_label=False,
|
|
|
|
| 762 |
elem_classes="input-textbox",
|
| 763 |
scale=1
|
| 764 |
)
|
| 765 |
+
|
| 766 |
with gr.Column(scale=1, min_width=70):
|
| 767 |
send = gr.Button(
|
| 768 |
"Send",
|
| 769 |
elem_classes="send-button custom-button",
|
| 770 |
scale=1
|
| 771 |
)
|
| 772 |
+
|
| 773 |
with gr.Column(scale=1, min_width=70):
|
| 774 |
clear = gr.Button(
|
| 775 |
"Clear",
|
| 776 |
elem_classes="clear-button custom-button",
|
| 777 |
scale=1
|
| 778 |
)
|
| 779 |
+
|
| 780 |
with gr.Accordion("๐ฎ Advanced Settings", open=False):
|
| 781 |
with gr.Row():
|
| 782 |
with gr.Column(scale=1):
|
|
|
|
| 817 |
current_file_context = None
|
| 818 |
return [], None, "Start a new conversation..."
|
| 819 |
|
| 820 |
+
# ์ด๋ฒคํธ ์ฐ๊ฒฐ
|
| 821 |
msg.submit(
|
| 822 |
stream_chat,
|
| 823 |
inputs=[msg, chatbot, file_upload, temperature, max_new_tokens, top_p, top_k, penalty],
|
|
|
|
| 841 |
queue=True
|
| 842 |
)
|
| 843 |
|
|
|
|
| 844 |
clear.click(
|
| 845 |
fn=clear_conversation,
|
| 846 |
outputs=[chatbot, file_upload, msg],
|
|
|
|
| 849 |
|
| 850 |
return demo
|
| 851 |
|
| 852 |
+
|
| 853 |
if __name__ == "__main__":
|
| 854 |
demo = create_demo()
|
| 855 |
+
demo.launch()
|