Spaces:
Sleeping
Sleeping
Commit
·
30be7bf
1
Parent(s):
f3e59e1
simplest version
Browse files- documents_prep.py +126 -164
- index_retriever.py +3 -3
- utils.py +155 -113
documents_prep.py
CHANGED
|
@@ -7,57 +7,71 @@ from llama_index.core.text_splitter import SentenceSplitter
|
|
| 7 |
from my_logging import log_message
|
| 8 |
|
| 9 |
# Configuration
|
| 10 |
-
CHUNK_SIZE =
|
| 11 |
CHUNK_OVERLAP = 256
|
|
|
|
| 12 |
def chunk_text_documents(documents):
|
| 13 |
-
"""Chunk with deduplication"""
|
| 14 |
text_splitter = SentenceSplitter(
|
| 15 |
chunk_size=CHUNK_SIZE,
|
| 16 |
-
chunk_overlap=
|
| 17 |
)
|
| 18 |
|
| 19 |
-
seen_texts = set()
|
| 20 |
chunked = []
|
| 21 |
-
|
| 22 |
for doc in documents:
|
| 23 |
-
text_normalized = doc.text.strip()
|
| 24 |
-
if len(text_normalized) < 50 or text_normalized in seen_texts:
|
| 25 |
-
continue
|
| 26 |
-
|
| 27 |
-
seen_texts.add(text_normalized)
|
| 28 |
-
|
| 29 |
chunks = text_splitter.get_nodes_from_documents([doc])
|
| 30 |
for i, chunk in enumerate(chunks):
|
| 31 |
chunk.metadata.update({
|
| 32 |
'chunk_id': i,
|
| 33 |
'total_chunks': len(chunks),
|
| 34 |
-
'chunk_size': len(chunk.text)
|
| 35 |
-
'document_group': normalize_doc_id(doc.metadata.get('document_id', 'unknown'))
|
| 36 |
})
|
| 37 |
chunked.append(chunk)
|
| 38 |
|
|
|
|
| 39 |
if chunked:
|
| 40 |
avg_size = sum(len(c.text) for c in chunked) / len(chunked)
|
| 41 |
-
|
|
|
|
|
|
|
|
|
|
| 42 |
|
| 43 |
return chunked
|
| 44 |
|
| 45 |
-
|
| 46 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 47 |
headers = table_data.get('headers', [])
|
| 48 |
rows = table_data.get('data', [])
|
| 49 |
table_num = table_data.get('table_number', 'unknown')
|
| 50 |
table_title = table_data.get('table_title', '')
|
| 51 |
section = table_data.get('section', '')
|
| 52 |
|
|
|
|
|
|
|
|
|
|
| 53 |
table_num_clean = str(table_num).strip()
|
| 54 |
|
| 55 |
-
# Create
|
| 56 |
import re
|
| 57 |
if 'приложени' in section.lower():
|
| 58 |
appendix_match = re.search(r'приложени[еия]\s*(\d+|[а-яА-Я])', section.lower())
|
| 59 |
if appendix_match:
|
| 60 |
-
|
|
|
|
| 61 |
else:
|
| 62 |
table_identifier = table_num_clean
|
| 63 |
else:
|
|
@@ -66,161 +80,128 @@ def chunk_table_by_rows(table_data, doc_id, max_chars=2000):
|
|
| 66 |
if not rows:
|
| 67 |
return []
|
| 68 |
|
| 69 |
-
|
| 70 |
-
base_content = f"Документ: {doc_id}\nТаблица: {table_identifier}\n"
|
| 71 |
-
if table_title:
|
| 72 |
-
base_content += f"Название: {table_title}\n"
|
| 73 |
-
if section:
|
| 74 |
-
base_content += f"Раздел: {section}\n"
|
| 75 |
-
|
| 76 |
-
header_content = ""
|
| 77 |
-
if headers:
|
| 78 |
-
header_content = "Столбцы: " + " | ".join(str(h) for h in headers) + "\n\n"
|
| 79 |
-
|
| 80 |
-
base_size = len(base_content) + len(header_content)
|
| 81 |
-
|
| 82 |
-
# Group rows by size
|
| 83 |
-
chunks = []
|
| 84 |
-
current_rows = []
|
| 85 |
-
current_size = base_size
|
| 86 |
|
| 87 |
-
|
| 88 |
-
|
| 89 |
-
|
| 90 |
-
row_str = " | ".join(f"{k}: {v}" for k, v in row.items()
|
| 91 |
-
if v and str(v).strip() and str(v).lower() not in ['nan', 'none', ''])
|
| 92 |
-
elif isinstance(row, list):
|
| 93 |
-
row_str = " | ".join(str(v) for v in row
|
| 94 |
-
if v and str(v).strip() and str(v).lower() not in ['nan', 'none', ''])
|
| 95 |
-
else:
|
| 96 |
-
row_str = str(row)
|
| 97 |
|
| 98 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 99 |
|
| 100 |
-
|
| 101 |
-
if current_size + row_size > max_chars and current_rows:
|
| 102 |
-
chunks.append(current_rows[:])
|
| 103 |
-
current_rows = []
|
| 104 |
-
current_size = base_size
|
| 105 |
|
| 106 |
-
|
| 107 |
-
current_size += row_size
|
| 108 |
|
| 109 |
-
|
| 110 |
-
|
| 111 |
-
chunks.append(current_rows)
|
| 112 |
|
| 113 |
-
|
| 114 |
-
|
| 115 |
-
|
| 116 |
-
|
| 117 |
-
content
|
| 118 |
-
|
| 119 |
-
|
| 120 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
| 121 |
|
| 122 |
-
|
| 123 |
-
if isinstance(row, dict):
|
| 124 |
-
parts = [f"{k}: {v}" for k, v in row.items()
|
| 125 |
-
if v and str(v).strip() and str(v).lower() not in ['nan', 'none', '']]
|
| 126 |
-
if parts:
|
| 127 |
-
content += f"{idx}. {' | '.join(parts)}\n"
|
| 128 |
-
elif isinstance(row, list):
|
| 129 |
-
parts = [str(v) for v in row if v and str(v).strip() and str(v).lower() not in ['nan', 'none', '']]
|
| 130 |
-
if parts:
|
| 131 |
-
content += f"{idx}. {' | '.join(parts)}\n"
|
| 132 |
|
| 133 |
metadata = {
|
| 134 |
'type': 'table',
|
| 135 |
'document_id': doc_id,
|
| 136 |
-
'document_group': normalize_doc_id(doc_id),
|
| 137 |
'table_number': table_num_clean,
|
| 138 |
'table_identifier': table_identifier,
|
| 139 |
'table_title': table_title,
|
| 140 |
'section': section,
|
| 141 |
-
'chunk_id':
|
| 142 |
-
'
|
| 143 |
-
'
|
| 144 |
-
'
|
|
|
|
|
|
|
|
|
|
| 145 |
}
|
| 146 |
|
| 147 |
-
|
| 148 |
|
| 149 |
-
|
| 150 |
-
log_message(f" Meta: doc={doc_id}, table={table_identifier}, group={metadata['document_group']}")
|
| 151 |
|
| 152 |
-
|
| 153 |
-
|
| 154 |
-
return documents
|
| 155 |
|
| 156 |
|
| 157 |
-
def
|
| 158 |
-
|
| 159 |
-
normalized = re.sub(r'\s+', ' ', str(doc_id).strip().upper())
|
| 160 |
-
normalized = normalized.replace('ГОСТ Р', 'ГОСТР').replace('ГОСТР', 'ГОСТ Р')
|
| 161 |
-
return normalized
|
| 162 |
-
|
| 163 |
-
|
| 164 |
-
def format_table_content(table_data, headers, rows, doc_id, table_identifier, chunk_info=""):
|
| 165 |
table_num = table_data.get('table_number', 'unknown')
|
| 166 |
table_title = table_data.get('table_title', '')
|
| 167 |
section = table_data.get('section', '')
|
| 168 |
|
| 169 |
-
#
|
| 170 |
content = f"ДОКУМЕНТ: {doc_id}\n"
|
| 171 |
content += f"ТАБЛИЦА: {table_identifier}\n"
|
| 172 |
-
|
| 173 |
-
|
| 174 |
-
doc_variations = [doc_id]
|
| 175 |
-
if 'Р' in doc_id:
|
| 176 |
-
doc_variations.append(doc_id.replace(' Р ', ' Р'))
|
| 177 |
-
doc_variations.append(doc_id.replace(' Р ', 'Р'))
|
| 178 |
-
|
| 179 |
-
for var in set(doc_variations):
|
| 180 |
-
content += f"ДОКУМЕНТ_ВАРИАНТ: {var}\n"
|
| 181 |
-
|
| 182 |
if table_title:
|
| 183 |
content += f"НАЗВАНИЕ: {table_title}\n"
|
| 184 |
if section:
|
| 185 |
content += f"РАЗДЕЛ: {section}\n"
|
| 186 |
-
|
| 187 |
content += f"{'='*70}\n\n"
|
| 188 |
|
| 189 |
-
# Enhanced search
|
| 190 |
-
content += f"
|
| 191 |
-
content += f"
|
| 192 |
-
content += f"
|
| 193 |
-
|
| 194 |
-
if table_title:
|
| 195 |
-
content += f"Название: {table_title}. "
|
| 196 |
|
| 197 |
if section:
|
| 198 |
-
content += f"
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 199 |
|
| 200 |
-
|
| 201 |
-
content += f"Таблицы документа {doc_id}. "
|
| 202 |
-
content += f"Содержание {doc_id}. "
|
| 203 |
|
| 204 |
if chunk_info:
|
| 205 |
-
content += f"{chunk_info}
|
| 206 |
|
| 207 |
-
content += f"\n\n
|
|
|
|
| 208 |
|
| 209 |
if headers:
|
| 210 |
-
|
|
|
|
| 211 |
|
|
|
|
| 212 |
for idx, row in enumerate(rows, 1):
|
| 213 |
if isinstance(row, dict):
|
| 214 |
parts = [f"{k}: {v}" for k, v in row.items()
|
| 215 |
-
if v and str(v).strip().lower() not in ['nan', 'none', ''
|
| 216 |
if parts:
|
| 217 |
content += f"{idx}. {' | '.join(parts)}\n"
|
| 218 |
elif isinstance(row, list):
|
| 219 |
-
parts = [str(v) for v in row
|
| 220 |
-
if v and str(v).strip().lower() not in ['nan', 'none', '', 'null']]
|
| 221 |
if parts:
|
| 222 |
content += f"{idx}. {' | '.join(parts)}\n"
|
| 223 |
|
|
|
|
|
|
|
|
|
|
| 224 |
return content
|
| 225 |
|
| 226 |
def load_json_documents(repo_id, hf_token, json_dir):
|
|
@@ -352,6 +333,7 @@ def load_json_documents(repo_id, hf_token, json_dir):
|
|
| 352 |
return documents
|
| 353 |
|
| 354 |
def extract_sections_from_json(json_path):
|
|
|
|
| 355 |
documents = []
|
| 356 |
|
| 357 |
try:
|
|
@@ -359,8 +341,8 @@ def extract_sections_from_json(json_path):
|
|
| 359 |
data = json.load(f)
|
| 360 |
|
| 361 |
doc_id = data.get('document_metadata', {}).get('document_id', 'unknown')
|
| 362 |
-
doc_id = normalize_doc_id(doc_id) # NORMALIZE
|
| 363 |
|
|
|
|
| 364 |
for section in data.get('sections', []):
|
| 365 |
if section.get('section_text', '').strip():
|
| 366 |
documents.append(Document(
|
|
@@ -368,11 +350,11 @@ def extract_sections_from_json(json_path):
|
|
| 368 |
metadata={
|
| 369 |
'type': 'text',
|
| 370 |
'document_id': doc_id,
|
| 371 |
-
'section_id': section.get('section_id', '')
|
| 372 |
-
'chunk_size': len(section['section_text'])
|
| 373 |
}
|
| 374 |
))
|
| 375 |
|
|
|
|
| 376 |
for subsection in section.get('subsections', []):
|
| 377 |
if subsection.get('subsection_text', '').strip():
|
| 378 |
documents.append(Document(
|
|
@@ -380,11 +362,11 @@ def extract_sections_from_json(json_path):
|
|
| 380 |
metadata={
|
| 381 |
'type': 'text',
|
| 382 |
'document_id': doc_id,
|
| 383 |
-
'section_id': subsection.get('subsection_id', '')
|
| 384 |
-
'chunk_size': len(subsection['subsection_text'])
|
| 385 |
}
|
| 386 |
))
|
| 387 |
|
|
|
|
| 388 |
for sub_sub in subsection.get('sub_subsections', []):
|
| 389 |
if sub_sub.get('sub_subsection_text', '').strip():
|
| 390 |
documents.append(Document(
|
|
@@ -392,8 +374,7 @@ def extract_sections_from_json(json_path):
|
|
| 392 |
metadata={
|
| 393 |
'type': 'text',
|
| 394 |
'document_id': doc_id,
|
| 395 |
-
'section_id': sub_sub.get('sub_subsection_id', '')
|
| 396 |
-
'chunk_size': len(sub_sub['sub_subsection_text'])
|
| 397 |
}
|
| 398 |
))
|
| 399 |
|
|
@@ -404,17 +385,13 @@ def extract_sections_from_json(json_path):
|
|
| 404 |
|
| 405 |
|
| 406 |
def load_table_documents(repo_id, hf_token, table_dir):
|
| 407 |
-
"""Load
|
| 408 |
log_message("Loading tables...")
|
| 409 |
|
| 410 |
files = list_repo_files(repo_id=repo_id, repo_type="dataset", token=hf_token)
|
| 411 |
table_files = [f for f in files if f.startswith(table_dir) and f.endswith('.json')]
|
| 412 |
|
| 413 |
-
log_message(f"Found {len(table_files)} table files")
|
| 414 |
-
|
| 415 |
all_chunks = []
|
| 416 |
-
doc_id_stats = {}
|
| 417 |
-
|
| 418 |
for file_path in table_files:
|
| 419 |
try:
|
| 420 |
local_path = hf_hub_download(
|
|
@@ -427,40 +404,32 @@ def load_table_documents(repo_id, hf_token, table_dir):
|
|
| 427 |
with open(local_path, 'r', encoding='utf-8') as f:
|
| 428 |
data = json.load(f)
|
| 429 |
|
|
|
|
| 430 |
file_doc_id = data.get('document_id', data.get('document', 'unknown'))
|
| 431 |
|
| 432 |
for sheet in data.get('sheets', []):
|
|
|
|
| 433 |
sheet_doc_id = sheet.get('document_id', sheet.get('document', file_doc_id))
|
| 434 |
|
| 435 |
-
#
|
| 436 |
-
if sheet_doc_id not in doc_id_stats:
|
| 437 |
-
doc_id_stats[sheet_doc_id] = 0
|
| 438 |
-
|
| 439 |
chunks = chunk_table_by_rows(sheet, sheet_doc_id)
|
| 440 |
all_chunks.extend(chunks)
|
| 441 |
-
doc_id_stats[sheet_doc_id] += len(chunks)
|
| 442 |
|
| 443 |
except Exception as e:
|
| 444 |
log_message(f"Error loading {file_path}: {e}")
|
| 445 |
|
| 446 |
-
|
| 447 |
-
log_message(f"\nTable loading summary:")
|
| 448 |
-
for doc_id, count in sorted(doc_id_stats.items()):
|
| 449 |
-
log_message(f" {doc_id}: {count} chunks")
|
| 450 |
-
|
| 451 |
-
log_message(f"\n✓ Total table chunks: {len(all_chunks)}")
|
| 452 |
return all_chunks
|
| 453 |
|
|
|
|
| 454 |
def load_image_documents(repo_id, hf_token, image_dir):
|
| 455 |
-
"""Load
|
| 456 |
log_message("Loading images...")
|
| 457 |
|
| 458 |
files = list_repo_files(repo_id=repo_id, repo_type="dataset", token=hf_token)
|
| 459 |
csv_files = [f for f in files if f.startswith(image_dir) and f.endswith('.csv')]
|
| 460 |
|
| 461 |
documents = []
|
| 462 |
-
seen = set()
|
| 463 |
-
|
| 464 |
for file_path in csv_files:
|
| 465 |
try:
|
| 466 |
local_path = hf_hub_download(
|
|
@@ -473,28 +442,22 @@ def load_image_documents(repo_id, hf_token, image_dir):
|
|
| 473 |
df = pd.read_csv(local_path)
|
| 474 |
|
| 475 |
for _, row in df.iterrows():
|
| 476 |
-
|
| 477 |
-
|
| 478 |
-
|
| 479 |
-
key = f"{doc_id}_{img_num}"
|
| 480 |
-
if key in seen:
|
| 481 |
-
continue
|
| 482 |
-
seen.add(key)
|
| 483 |
-
|
| 484 |
-
content = f"Документ: {doc_id}\n"
|
| 485 |
-
content += f"Рисунок: {img_num}\n"
|
| 486 |
content += f"Название: {row.get('Название изображения', '')}\n"
|
| 487 |
content += f"Описание: {row.get('Описание изображение', '')}\n"
|
|
|
|
|
|
|
|
|
|
| 488 |
|
| 489 |
documents.append(Document(
|
| 490 |
text=content,
|
| 491 |
metadata={
|
| 492 |
'type': 'image',
|
| 493 |
-
'document_id':
|
| 494 |
-
'
|
| 495 |
-
'image_number': img_num,
|
| 496 |
'section': str(row.get('Раздел документа', '')),
|
| 497 |
-
'chunk_size':
|
| 498 |
}
|
| 499 |
))
|
| 500 |
except Exception as e:
|
|
@@ -502,12 +465,11 @@ def load_image_documents(repo_id, hf_token, image_dir):
|
|
| 502 |
|
| 503 |
if documents:
|
| 504 |
avg_size = sum(d.metadata['chunk_size'] for d in documents) / len(documents)
|
| 505 |
-
log_message(f"✓
|
| 506 |
|
| 507 |
return documents
|
| 508 |
|
| 509 |
|
| 510 |
-
|
| 511 |
def load_all_documents(repo_id, hf_token, json_dir, table_dir, image_dir):
|
| 512 |
"""Main loader - combines all document types"""
|
| 513 |
log_message("="*60)
|
|
|
|
| 7 |
from my_logging import log_message
|
| 8 |
|
| 9 |
# Configuration
|
| 10 |
+
CHUNK_SIZE = 1024
|
| 11 |
CHUNK_OVERLAP = 256
|
| 12 |
+
|
| 13 |
def chunk_text_documents(documents):
|
|
|
|
| 14 |
text_splitter = SentenceSplitter(
|
| 15 |
chunk_size=CHUNK_SIZE,
|
| 16 |
+
chunk_overlap=CHUNK_OVERLAP
|
| 17 |
)
|
| 18 |
|
|
|
|
| 19 |
chunked = []
|
|
|
|
| 20 |
for doc in documents:
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 21 |
chunks = text_splitter.get_nodes_from_documents([doc])
|
| 22 |
for i, chunk in enumerate(chunks):
|
| 23 |
chunk.metadata.update({
|
| 24 |
'chunk_id': i,
|
| 25 |
'total_chunks': len(chunks),
|
| 26 |
+
'chunk_size': len(chunk.text) # Add chunk size
|
|
|
|
| 27 |
})
|
| 28 |
chunked.append(chunk)
|
| 29 |
|
| 30 |
+
# Log statistics
|
| 31 |
if chunked:
|
| 32 |
avg_size = sum(len(c.text) for c in chunked) / len(chunked)
|
| 33 |
+
min_size = min(len(c.text) for c in chunked)
|
| 34 |
+
max_size = max(len(c.text) for c in chunked)
|
| 35 |
+
log_message(f"✓ Text: {len(documents)} docs → {len(chunked)} chunks")
|
| 36 |
+
log_message(f" Size stats: avg={avg_size:.0f}, min={min_size}, max={max_size} chars")
|
| 37 |
|
| 38 |
return chunked
|
| 39 |
|
| 40 |
+
|
| 41 |
+
def normalize_doc_id(doc_id):
|
| 42 |
+
"""Normalize document ID for consistent matching"""
|
| 43 |
+
if not doc_id or doc_id == 'unknown':
|
| 44 |
+
return doc_id
|
| 45 |
+
|
| 46 |
+
doc_id = str(doc_id).strip()
|
| 47 |
+
|
| 48 |
+
# Normalize spacing: "ГОСТ Р" variations
|
| 49 |
+
import re
|
| 50 |
+
doc_id = re.sub(r'ГОСТ\s*Р', 'ГОСТ Р', doc_id, flags=re.IGNORECASE)
|
| 51 |
+
doc_id = re.sub(r'НП\s*-', 'НП-', doc_id, flags=re.IGNORECASE)
|
| 52 |
+
|
| 53 |
+
return doc_id
|
| 54 |
+
|
| 55 |
+
|
| 56 |
+
def chunk_table_by_rows(table_data, doc_id, max_rows=2):
|
| 57 |
headers = table_data.get('headers', [])
|
| 58 |
rows = table_data.get('data', [])
|
| 59 |
table_num = table_data.get('table_number', 'unknown')
|
| 60 |
table_title = table_data.get('table_title', '')
|
| 61 |
section = table_data.get('section', '')
|
| 62 |
|
| 63 |
+
# NORMALIZE document ID
|
| 64 |
+
doc_id = normalize_doc_id(doc_id)
|
| 65 |
+
|
| 66 |
table_num_clean = str(table_num).strip()
|
| 67 |
|
| 68 |
+
# Create section-aware identifier
|
| 69 |
import re
|
| 70 |
if 'приложени' in section.lower():
|
| 71 |
appendix_match = re.search(r'приложени[еия]\s*(\d+|[а-яА-Я])', section.lower())
|
| 72 |
if appendix_match:
|
| 73 |
+
appendix_num = appendix_match.group(1).upper()
|
| 74 |
+
table_identifier = f"{table_num_clean} Приложение {appendix_num}"
|
| 75 |
else:
|
| 76 |
table_identifier = table_num_clean
|
| 77 |
else:
|
|
|
|
| 80 |
if not rows:
|
| 81 |
return []
|
| 82 |
|
| 83 |
+
log_message(f" 📊 Processing: {doc_id} - {table_identifier} ({len(rows)} rows)")
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 84 |
|
| 85 |
+
if len(rows) <= max_rows:
|
| 86 |
+
content = format_table_content(table_data, headers, rows, doc_id, table_identifier)
|
| 87 |
+
chunk_size = len(content)
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 88 |
|
| 89 |
+
metadata = {
|
| 90 |
+
'type': 'table',
|
| 91 |
+
'document_id': doc_id,
|
| 92 |
+
'table_number': table_num_clean,
|
| 93 |
+
'table_identifier': table_identifier,
|
| 94 |
+
'table_title': table_title,
|
| 95 |
+
'section': section,
|
| 96 |
+
'total_rows': len(rows),
|
| 97 |
+
'chunk_size': chunk_size,
|
| 98 |
+
'is_complete_table': True
|
| 99 |
+
}
|
| 100 |
|
| 101 |
+
log_message(f" Chunk: 1/1, {chunk_size} chars, doc={doc_id}, table={table_identifier}")
|
|
|
|
|
|
|
|
|
|
|
|
|
| 102 |
|
| 103 |
+
return [Document(text=content, metadata=metadata)]
|
|
|
|
| 104 |
|
| 105 |
+
chunks = []
|
| 106 |
+
overlap = 1
|
|
|
|
| 107 |
|
| 108 |
+
for i in range(0, len(rows), max_rows - overlap):
|
| 109 |
+
chunk_rows = rows[i:min(i+max_rows, len(rows))]
|
| 110 |
+
chunk_num = i // (max_rows - overlap)
|
| 111 |
+
|
| 112 |
+
content = format_table_content(
|
| 113 |
+
table_data,
|
| 114 |
+
headers,
|
| 115 |
+
chunk_rows,
|
| 116 |
+
doc_id,
|
| 117 |
+
table_identifier,
|
| 118 |
+
chunk_info=f"Строки {i+1}-{i+len(chunk_rows)} из {len(rows)}"
|
| 119 |
+
)
|
| 120 |
|
| 121 |
+
chunk_size = len(content)
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 122 |
|
| 123 |
metadata = {
|
| 124 |
'type': 'table',
|
| 125 |
'document_id': doc_id,
|
|
|
|
| 126 |
'table_number': table_num_clean,
|
| 127 |
'table_identifier': table_identifier,
|
| 128 |
'table_title': table_title,
|
| 129 |
'section': section,
|
| 130 |
+
'chunk_id': chunk_num,
|
| 131 |
+
'row_start': i,
|
| 132 |
+
'row_end': i + len(chunk_rows),
|
| 133 |
+
'total_rows': len(rows),
|
| 134 |
+
'chunk_size': chunk_size,
|
| 135 |
+
'total_chunks': (len(rows) + max_rows - overlap - 1) // (max_rows - overlap),
|
| 136 |
+
'is_complete_table': False
|
| 137 |
}
|
| 138 |
|
| 139 |
+
log_message(f" Chunk: {chunk_num+1}, rows {i}-{i+len(chunk_rows)}, {chunk_size} chars")
|
| 140 |
|
| 141 |
+
chunks.append(Document(text=content, metadata=metadata))
|
|
|
|
| 142 |
|
| 143 |
+
return chunks
|
|
|
|
|
|
|
| 144 |
|
| 145 |
|
| 146 |
+
def format_table_content(table_data, headers, rows, table_identifier, chunk_info=""):
|
| 147 |
+
doc_id = table_data.get('document_id', table_data.get('document', 'unknown'))
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 148 |
table_num = table_data.get('table_number', 'unknown')
|
| 149 |
table_title = table_data.get('table_title', '')
|
| 150 |
section = table_data.get('section', '')
|
| 151 |
|
| 152 |
+
# Use enhanced identifier
|
| 153 |
content = f"ДОКУМЕНТ: {doc_id}\n"
|
| 154 |
content += f"ТАБЛИЦА: {table_identifier}\n"
|
| 155 |
+
content += f"ПОЛНОЕ НАЗВАНИЕ: {table_identifier}\n"
|
| 156 |
+
content += f"НОМЕР ТАБЛИЦЫ: {table_num}\n"
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 157 |
if table_title:
|
| 158 |
content += f"НАЗВАНИЕ: {table_title}\n"
|
| 159 |
if section:
|
| 160 |
content += f"РАЗДЕЛ: {section}\n"
|
|
|
|
| 161 |
content += f"{'='*70}\n\n"
|
| 162 |
|
| 163 |
+
# Enhanced search keywords
|
| 164 |
+
content += f"Это таблица {table_identifier} из документа {doc_id}. "
|
| 165 |
+
content += f"Идентификатор таблицы: {table_identifier}. "
|
| 166 |
+
content += f"Номер: {table_num}. "
|
| 167 |
+
content += f"Документ: {doc_id}. "
|
|
|
|
|
|
|
| 168 |
|
| 169 |
if section:
|
| 170 |
+
content += f"Находится в разделе: {section}. "
|
| 171 |
+
if 'приложени' in section.lower():
|
| 172 |
+
content += f"Таблица из приложения. "
|
| 173 |
+
|
| 174 |
+
if table_title:
|
| 175 |
+
content += f"Название таблицы: {table_title}. "
|
| 176 |
+
content += f"Таблица о: {table_title}. "
|
| 177 |
|
| 178 |
+
content += f"Поиск: таблица {table_identifier} {doc_id}. "
|
|
|
|
|
|
|
| 179 |
|
| 180 |
if chunk_info:
|
| 181 |
+
content += f"\n{chunk_info}\n"
|
| 182 |
|
| 183 |
+
content += f"\n\nСОДЕРЖИМОЕ ТАБЛИЦЫ {table_identifier}:\n"
|
| 184 |
+
content += f"="*70 + "\n\n"
|
| 185 |
|
| 186 |
if headers:
|
| 187 |
+
header_str = ' | '.join(str(h) for h in headers)
|
| 188 |
+
content += f"ЗАГОЛОВКИ СТОЛБЦОВ:\n{header_str}\n\n"
|
| 189 |
|
| 190 |
+
content += f"ДАННЫЕ ТАБЛИЦЫ:\n"
|
| 191 |
for idx, row in enumerate(rows, 1):
|
| 192 |
if isinstance(row, dict):
|
| 193 |
parts = [f"{k}: {v}" for k, v in row.items()
|
| 194 |
+
if v and str(v).strip() and str(v).lower() not in ['nan', 'none', '']]
|
| 195 |
if parts:
|
| 196 |
content += f"{idx}. {' | '.join(parts)}\n"
|
| 197 |
elif isinstance(row, list):
|
| 198 |
+
parts = [str(v) for v in row if v and str(v).strip() and str(v).lower() not in ['nan', 'none', '']]
|
|
|
|
| 199 |
if parts:
|
| 200 |
content += f"{idx}. {' | '.join(parts)}\n"
|
| 201 |
|
| 202 |
+
content += f"\n{'='*70}\n"
|
| 203 |
+
content += f"КОНЕЦ ТАБЛИЦЫ {table_identifier} ИЗ {doc_id}\n"
|
| 204 |
+
|
| 205 |
return content
|
| 206 |
|
| 207 |
def load_json_documents(repo_id, hf_token, json_dir):
|
|
|
|
| 333 |
return documents
|
| 334 |
|
| 335 |
def extract_sections_from_json(json_path):
|
| 336 |
+
"""Extract sections from a single JSON file"""
|
| 337 |
documents = []
|
| 338 |
|
| 339 |
try:
|
|
|
|
| 341 |
data = json.load(f)
|
| 342 |
|
| 343 |
doc_id = data.get('document_metadata', {}).get('document_id', 'unknown')
|
|
|
|
| 344 |
|
| 345 |
+
# Extract all section levels
|
| 346 |
for section in data.get('sections', []):
|
| 347 |
if section.get('section_text', '').strip():
|
| 348 |
documents.append(Document(
|
|
|
|
| 350 |
metadata={
|
| 351 |
'type': 'text',
|
| 352 |
'document_id': doc_id,
|
| 353 |
+
'section_id': section.get('section_id', '')
|
|
|
|
| 354 |
}
|
| 355 |
))
|
| 356 |
|
| 357 |
+
# Subsections
|
| 358 |
for subsection in section.get('subsections', []):
|
| 359 |
if subsection.get('subsection_text', '').strip():
|
| 360 |
documents.append(Document(
|
|
|
|
| 362 |
metadata={
|
| 363 |
'type': 'text',
|
| 364 |
'document_id': doc_id,
|
| 365 |
+
'section_id': subsection.get('subsection_id', '')
|
|
|
|
| 366 |
}
|
| 367 |
))
|
| 368 |
|
| 369 |
+
# Sub-subsections
|
| 370 |
for sub_sub in subsection.get('sub_subsections', []):
|
| 371 |
if sub_sub.get('sub_subsection_text', '').strip():
|
| 372 |
documents.append(Document(
|
|
|
|
| 374 |
metadata={
|
| 375 |
'type': 'text',
|
| 376 |
'document_id': doc_id,
|
| 377 |
+
'section_id': sub_sub.get('sub_subsection_id', '')
|
|
|
|
| 378 |
}
|
| 379 |
))
|
| 380 |
|
|
|
|
| 385 |
|
| 386 |
|
| 387 |
def load_table_documents(repo_id, hf_token, table_dir):
|
| 388 |
+
"""Load and chunk tables"""
|
| 389 |
log_message("Loading tables...")
|
| 390 |
|
| 391 |
files = list_repo_files(repo_id=repo_id, repo_type="dataset", token=hf_token)
|
| 392 |
table_files = [f for f in files if f.startswith(table_dir) and f.endswith('.json')]
|
| 393 |
|
|
|
|
|
|
|
| 394 |
all_chunks = []
|
|
|
|
|
|
|
| 395 |
for file_path in table_files:
|
| 396 |
try:
|
| 397 |
local_path = hf_hub_download(
|
|
|
|
| 404 |
with open(local_path, 'r', encoding='utf-8') as f:
|
| 405 |
data = json.load(f)
|
| 406 |
|
| 407 |
+
# Extract file-level document_id
|
| 408 |
file_doc_id = data.get('document_id', data.get('document', 'unknown'))
|
| 409 |
|
| 410 |
for sheet in data.get('sheets', []):
|
| 411 |
+
# Use sheet-level document_id if available, otherwise use file-level
|
| 412 |
sheet_doc_id = sheet.get('document_id', sheet.get('document', file_doc_id))
|
| 413 |
|
| 414 |
+
# CRITICAL: Pass document_id to chunk function
|
|
|
|
|
|
|
|
|
|
| 415 |
chunks = chunk_table_by_rows(sheet, sheet_doc_id)
|
| 416 |
all_chunks.extend(chunks)
|
|
|
|
| 417 |
|
| 418 |
except Exception as e:
|
| 419 |
log_message(f"Error loading {file_path}: {e}")
|
| 420 |
|
| 421 |
+
log_message(f"✓ Loaded {len(all_chunks)} table chunks")
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 422 |
return all_chunks
|
| 423 |
|
| 424 |
+
|
| 425 |
def load_image_documents(repo_id, hf_token, image_dir):
|
| 426 |
+
"""Load image descriptions"""
|
| 427 |
log_message("Loading images...")
|
| 428 |
|
| 429 |
files = list_repo_files(repo_id=repo_id, repo_type="dataset", token=hf_token)
|
| 430 |
csv_files = [f for f in files if f.startswith(image_dir) and f.endswith('.csv')]
|
| 431 |
|
| 432 |
documents = []
|
|
|
|
|
|
|
| 433 |
for file_path in csv_files:
|
| 434 |
try:
|
| 435 |
local_path = hf_hub_download(
|
|
|
|
| 442 |
df = pd.read_csv(local_path)
|
| 443 |
|
| 444 |
for _, row in df.iterrows():
|
| 445 |
+
content = f"Документ: {row.get('Обозначение документа', 'unknown')}\n"
|
| 446 |
+
content += f"Рисунок: {row.get('№ Изображения', 'unknown')}\n"
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 447 |
content += f"Название: {row.get('Название изображения', '')}\n"
|
| 448 |
content += f"Описание: {row.get('Описание изображение', '')}\n"
|
| 449 |
+
content += f"Раздел: {row.get('Раздел документа', '')}\n"
|
| 450 |
+
|
| 451 |
+
chunk_size = len(content)
|
| 452 |
|
| 453 |
documents.append(Document(
|
| 454 |
text=content,
|
| 455 |
metadata={
|
| 456 |
'type': 'image',
|
| 457 |
+
'document_id': str(row.get('Обозначение документа', 'unknown')),
|
| 458 |
+
'image_number': str(row.get('№ Изображения', 'unknown')),
|
|
|
|
| 459 |
'section': str(row.get('Раздел документа', '')),
|
| 460 |
+
'chunk_size': chunk_size
|
| 461 |
}
|
| 462 |
))
|
| 463 |
except Exception as e:
|
|
|
|
| 465 |
|
| 466 |
if documents:
|
| 467 |
avg_size = sum(d.metadata['chunk_size'] for d in documents) / len(documents)
|
| 468 |
+
log_message(f"✓ Loaded {len(documents)} images (avg size: {avg_size:.0f} chars)")
|
| 469 |
|
| 470 |
return documents
|
| 471 |
|
| 472 |
|
|
|
|
| 473 |
def load_all_documents(repo_id, hf_token, json_dir, table_dir, image_dir):
|
| 474 |
"""Main loader - combines all document types"""
|
| 475 |
log_message("="*60)
|
index_retriever.py
CHANGED
|
@@ -35,19 +35,19 @@ def create_query_engine(vector_index):
|
|
| 35 |
# Vector retriever
|
| 36 |
vector_retriever = VectorIndexRetriever(
|
| 37 |
index=vector_index,
|
| 38 |
-
similarity_top_k=
|
| 39 |
)
|
| 40 |
|
| 41 |
# BM25 retriever
|
| 42 |
bm25_retriever = BM25Retriever.from_defaults(
|
| 43 |
docstore=vector_index.docstore,
|
| 44 |
-
similarity_top_k=
|
| 45 |
)
|
| 46 |
|
| 47 |
# Hybrid fusion
|
| 48 |
hybrid_retriever = QueryFusionRetriever(
|
| 49 |
[vector_retriever, bm25_retriever],
|
| 50 |
-
similarity_top_k=
|
| 51 |
num_queries=1
|
| 52 |
)
|
| 53 |
|
|
|
|
| 35 |
# Vector retriever
|
| 36 |
vector_retriever = VectorIndexRetriever(
|
| 37 |
index=vector_index,
|
| 38 |
+
similarity_top_k=50
|
| 39 |
)
|
| 40 |
|
| 41 |
# BM25 retriever
|
| 42 |
bm25_retriever = BM25Retriever.from_defaults(
|
| 43 |
docstore=vector_index.docstore,
|
| 44 |
+
similarity_top_k=50
|
| 45 |
)
|
| 46 |
|
| 47 |
# Hybrid fusion
|
| 48 |
hybrid_retriever = QueryFusionRetriever(
|
| 49 |
[vector_retriever, bm25_retriever],
|
| 50 |
+
similarity_top_k=60,
|
| 51 |
num_queries=1
|
| 52 |
)
|
| 53 |
|
utils.py
CHANGED
|
@@ -40,96 +40,74 @@ def preprocess_query(question):
|
|
| 40 |
import re
|
| 41 |
|
| 42 |
question_lower = question.lower()
|
|
|
|
|
|
|
|
|
|
|
|
|
| 43 |
enhanced_query = question
|
| 44 |
|
| 45 |
-
# Detect "list all tables" queries - handle differently
|
| 46 |
-
if any(phrase in question_lower for phrase in ['какие таблиц', 'список таблиц', 'перечисл', 'все таблиц']):
|
| 47 |
-
# For listing queries, just extract document ID
|
| 48 |
-
doc_match = re.search(r'(гост|нп|му)[^\s]*\s*р?\s*[№-]*\s*([0-9\.-]+)', question_lower)
|
| 49 |
-
if doc_match:
|
| 50 |
-
doc_id = f"{doc_match.group(1).upper()} Р {doc_match.group(2)}"
|
| 51 |
-
enhanced_query = f"документ {doc_id} таблица"
|
| 52 |
-
return enhanced_query
|
| 53 |
-
|
| 54 |
-
# For specific table queries
|
| 55 |
-
table_match = re.search(r'табли[цу]\w*\s+(?:№|номер)?\s*([а-яa-z0-9\.]+)', question_lower)
|
| 56 |
-
if table_match:
|
| 57 |
-
table_num = table_match.group(1).upper()
|
| 58 |
-
enhanced_query += f" таблица {table_num}"
|
| 59 |
-
|
| 60 |
-
# Document detection
|
| 61 |
-
doc_match = re.search(r'(гост|нп|му)[^\s]*\s*р?\s*[№-]*\s*([0-9\.-]+)', question_lower)
|
| 62 |
if doc_match:
|
| 63 |
-
|
| 64 |
-
|
|
|
|
|
|
|
|
|
|
| 65 |
|
| 66 |
return enhanced_query
|
| 67 |
|
| 68 |
def answer_question(question, query_engine, reranker):
|
| 69 |
try:
|
| 70 |
-
log_message(f"
|
| 71 |
-
log_message(f"QUERY: {question}")
|
| 72 |
|
| 73 |
enhanced_query = preprocess_query(question)
|
| 74 |
-
|
| 75 |
-
|
| 76 |
-
# Detect listing queries - need MORE chunks
|
| 77 |
-
is_listing_query = any(phrase in question.lower()
|
| 78 |
-
for phrase in ['какие таблиц', 'список', 'перечисл', 'все таблиц'])
|
| 79 |
|
| 80 |
retrieved = query_engine.retriever.retrieve(enhanced_query)
|
| 81 |
-
log_message(f"
|
| 82 |
|
| 83 |
-
# Log retrieved docs
|
| 84 |
doc_stats = {}
|
| 85 |
for n in retrieved:
|
| 86 |
doc_id = n.metadata.get('document_id', 'unknown')
|
| 87 |
-
|
| 88 |
|
| 89 |
-
if
|
| 90 |
-
doc_stats[
|
| 91 |
|
| 92 |
-
if
|
| 93 |
table_id = n.metadata.get('table_identifier', n.metadata.get('table_number', '?'))
|
| 94 |
-
doc_stats[
|
|
|
|
|
|
|
| 95 |
else:
|
| 96 |
-
doc_stats[
|
| 97 |
|
| 98 |
for doc_id in sorted(doc_stats.keys()):
|
| 99 |
stats = doc_stats[doc_id]
|
| 100 |
-
|
| 101 |
if stats['tables']:
|
| 102 |
-
|
| 103 |
-
|
| 104 |
-
|
| 105 |
-
|
| 106 |
-
|
| 107 |
-
|
| 108 |
-
|
| 109 |
-
|
| 110 |
-
|
|
|
|
|
|
|
|
|
|
| 111 |
|
| 112 |
-
|
| 113 |
-
|
| 114 |
-
for n in reranked:
|
| 115 |
-
doc_group = n.metadata.get('document_group', n.metadata.get('document_id', 'unknown'))
|
| 116 |
-
|
| 117 |
-
if doc_group not in doc_stats_reranked:
|
| 118 |
-
doc_stats_reranked[doc_group] = {'tables': set(), 'text': 0}
|
| 119 |
-
|
| 120 |
-
if n.metadata.get('type') == 'table':
|
| 121 |
-
table_id = n.metadata.get('table_identifier', n.metadata.get('table_number', '?'))
|
| 122 |
-
doc_stats_reranked[doc_group]['tables'].add(table_id)
|
| 123 |
-
else:
|
| 124 |
-
doc_stats_reranked[doc_group]['text'] += 1
|
| 125 |
|
| 126 |
-
for
|
| 127 |
-
|
| 128 |
-
|
| 129 |
-
|
| 130 |
-
|
| 131 |
|
| 132 |
-
# Build context
|
| 133 |
context_parts = []
|
| 134 |
for n in reranked:
|
| 135 |
meta = n.metadata
|
|
@@ -137,48 +115,103 @@ def answer_question(question, query_engine, reranker):
|
|
| 137 |
doc_type = meta.get('type', 'text')
|
| 138 |
|
| 139 |
if doc_type == 'table':
|
| 140 |
-
|
| 141 |
title = meta.get('table_title', '')
|
| 142 |
-
source_label = f"[{
|
| 143 |
if title:
|
| 144 |
source_label += f" {title}"
|
|
|
|
|
|
|
|
|
|
| 145 |
else:
|
| 146 |
-
|
|
|
|
| 147 |
|
| 148 |
-
context_parts.append(f"{source_label}\n{n.text
|
| 149 |
-
|
| 150 |
-
context = "\n\n" + ("="*50 + "\n\n").join(context_parts)
|
| 151 |
|
| 152 |
-
|
| 153 |
-
|
| 154 |
-
|
|
|
|
| 155 |
|
| 156 |
-
|
| 157 |
-
{context}
|
| 158 |
|
| 159 |
-
|
|
|
|
| 160 |
|
| 161 |
-
|
| 162 |
-
1. Перечисли ВСЕ таблицы, найденные в контексте для запрошенного документа
|
| 163 |
-
2. Укажи номер таблицы и название (если есть)
|
| 164 |
-
3. Если таблиц нет - скажи прямо
|
| 165 |
|
| 166 |
-
ОТВЕТ (
|
| 167 |
-
|
| 168 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 169 |
|
| 170 |
КОНТЕКСТ:
|
| 171 |
{context}
|
| 172 |
|
| 173 |
ВОПРОС: {question}
|
| 174 |
-
|
| 175 |
-
ИНСТРУКЦИИ:
|
| 176 |
-
1. Отвечай ТОЛЬКО на основе контекста
|
| 177 |
-
2. Укажи источник (документ, таблицу)
|
| 178 |
-
3. Если нужно показать содержимое таблицы - покажи ВСЕ данные
|
| 179 |
-
4. Если информации нет - скажи прямо
|
| 180 |
-
|
| 181 |
-
ОТВЕТ:"""
|
| 182 |
|
| 183 |
response = query_engine.query(prompt)
|
| 184 |
sources = format_sources(reranked)
|
|
@@ -190,45 +223,54 @@ def answer_question(question, query_engine, reranker):
|
|
| 190 |
import traceback
|
| 191 |
log_message(traceback.format_exc())
|
| 192 |
return f"Ошибка: {e}", ""
|
| 193 |
-
|
| 194 |
-
|
|
|
|
| 195 |
if not nodes:
|
| 196 |
return []
|
| 197 |
|
|
|
|
| 198 |
pairs = [[query, n.text] for n in nodes]
|
| 199 |
scores = reranker.predict(pairs)
|
| 200 |
|
|
|
|
| 201 |
scored = sorted(zip(nodes, scores), key=lambda x: x[1], reverse=True)
|
| 202 |
|
| 203 |
-
|
| 204 |
-
|
| 205 |
-
# More lenient filtering
|
| 206 |
filtered = [(n, s) for n, s in scored if s >= min_score]
|
| 207 |
|
| 208 |
if not filtered:
|
| 209 |
-
|
|
|
|
| 210 |
filtered = [(n, s) for n, s in scored if s >= cutoff][:top_k]
|
| 211 |
|
| 212 |
-
#
|
| 213 |
-
|
| 214 |
-
for node, score in filtered:
|
| 215 |
-
doc_group = node.metadata.get('document_group', node.metadata.get('document_id', 'unknown'))
|
| 216 |
-
if doc_group not in doc_groups:
|
| 217 |
-
doc_groups[doc_group] = []
|
| 218 |
-
doc_groups[doc_group].append((node, score))
|
| 219 |
|
| 220 |
-
#
|
| 221 |
selected = []
|
| 222 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 223 |
|
| 224 |
-
|
| 225 |
-
|
|
|
|
|
|
|
|
|
|
| 226 |
|
| 227 |
-
|
| 228 |
-
|
| 229 |
-
|
| 230 |
-
selected.
|
|
|
|
| 231 |
|
| 232 |
-
log_message(f"Reranked: {len(filtered)} → {len(selected)} (from {len(
|
| 233 |
|
| 234 |
-
return selected
|
|
|
|
| 40 |
import re
|
| 41 |
|
| 42 |
question_lower = question.lower()
|
| 43 |
+
|
| 44 |
+
# Extract document ID and normalize
|
| 45 |
+
doc_match = re.search(r'(гост|нп|му)\s*р?\s*[№-]*\s*([0-9\.-]+)', question_lower)
|
| 46 |
+
|
| 47 |
enhanced_query = question
|
| 48 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 49 |
if doc_match:
|
| 50 |
+
doc_type = doc_match.group(1).upper()
|
| 51 |
+
doc_num = doc_match.group(2)
|
| 52 |
+
|
| 53 |
+
# Add normalized versions
|
| 54 |
+
enhanced_query += f" {doc_type} Р {doc_num}"
|
| 55 |
|
| 56 |
return enhanced_query
|
| 57 |
|
| 58 |
def answer_question(question, query_engine, reranker):
|
| 59 |
try:
|
| 60 |
+
log_message(f"Query: {question}")
|
|
|
|
| 61 |
|
| 62 |
enhanced_query = preprocess_query(question)
|
| 63 |
+
if enhanced_query != question:
|
| 64 |
+
log_message(f"Enhanced query: {enhanced_query}")
|
|
|
|
|
|
|
|
|
|
| 65 |
|
| 66 |
retrieved = query_engine.retriever.retrieve(enhanced_query)
|
| 67 |
+
log_message(f"Retrieved {len(retrieved)} nodes")
|
| 68 |
|
|
|
|
| 69 |
doc_stats = {}
|
| 70 |
for n in retrieved:
|
| 71 |
doc_id = n.metadata.get('document_id', 'unknown')
|
| 72 |
+
doc_type = n.metadata.get('type', 'text')
|
| 73 |
|
| 74 |
+
if doc_id not in doc_stats:
|
| 75 |
+
doc_stats[doc_id] = {'tables': set(), 'text': 0, 'images': 0}
|
| 76 |
|
| 77 |
+
if doc_type == 'table':
|
| 78 |
table_id = n.metadata.get('table_identifier', n.metadata.get('table_number', '?'))
|
| 79 |
+
doc_stats[doc_id]['tables'].add(table_id)
|
| 80 |
+
elif doc_type == 'image':
|
| 81 |
+
doc_stats[doc_id]['images'] += 1
|
| 82 |
else:
|
| 83 |
+
doc_stats[doc_id]['text'] += 1
|
| 84 |
|
| 85 |
for doc_id in sorted(doc_stats.keys()):
|
| 86 |
stats = doc_stats[doc_id]
|
| 87 |
+
parts = []
|
| 88 |
if stats['tables']:
|
| 89 |
+
parts.append(f"tables={list(stats['tables'])[:5]}")
|
| 90 |
+
if stats['text']:
|
| 91 |
+
parts.append(f"text={stats['text']}")
|
| 92 |
+
if stats['images']:
|
| 93 |
+
parts.append(f"images={stats['images']}")
|
| 94 |
+
log_message(f" {doc_id}: {', '.join(parts)}")
|
| 95 |
+
|
| 96 |
+
doc_ids = [n.metadata.get('document_id', 'unknown') for n in retrieved]
|
| 97 |
+
table_nums = [n.metadata.get('table_number', '') for n in retrieved if n.metadata.get('type') == 'table']
|
| 98 |
+
log_message(f"Retrieved from documents: {set(doc_ids)}")
|
| 99 |
+
if table_nums:
|
| 100 |
+
log_message(f"Retrieved tables: {set(table_nums)}")
|
| 101 |
|
| 102 |
+
reranked = rerank_nodes(question, retrieved, reranker, top_k=25)
|
| 103 |
+
log_message(f"Reranked to {len(reranked)} nodes")
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 104 |
|
| 105 |
+
doc_ids_reranked = [n.metadata.get('document_id', 'unknown') for n in reranked]
|
| 106 |
+
table_nums_reranked = [n.metadata.get('table_number', '') for n in reranked if n.metadata.get('type') == 'table']
|
| 107 |
+
log_message(f"After reranking - documents: {set(doc_ids_reranked)}")
|
| 108 |
+
if table_nums_reranked:
|
| 109 |
+
log_message(f"After reranking - tables: {set(table_nums_reranked)}")
|
| 110 |
|
|
|
|
| 111 |
context_parts = []
|
| 112 |
for n in reranked:
|
| 113 |
meta = n.metadata
|
|
|
|
| 115 |
doc_type = meta.get('type', 'text')
|
| 116 |
|
| 117 |
if doc_type == 'table':
|
| 118 |
+
table_num = meta.get('table_number', 'unknown')
|
| 119 |
title = meta.get('table_title', '')
|
| 120 |
+
source_label = f"[ТАБЛИЦА {table_num} - {doc_id}]"
|
| 121 |
if title:
|
| 122 |
source_label += f" {title}"
|
| 123 |
+
elif doc_type == 'image':
|
| 124 |
+
img_num = meta.get('image_number', 'unknown')
|
| 125 |
+
source_label = f"[РИСУНОК {img_num} - {doc_id}]"
|
| 126 |
else:
|
| 127 |
+
section = meta.get('section_id', '')
|
| 128 |
+
source_label = f"[{doc_id} - {section}]"
|
| 129 |
|
| 130 |
+
context_parts.append(f"{source_label}\n{n.text}")
|
|
|
|
|
|
|
| 131 |
|
| 132 |
+
context = "\n\n" + ("="*70 + "\n\n").join(context_parts)
|
| 133 |
+
from config import CUSTOM_PROMPT
|
| 134 |
+
prompt = f"""
|
| 135 |
+
Вы являетесь высокоспециализированным Ассистентом для анализа нормативных документов (AIEXP). Ваша цель - предоставлять точные, корректные и контекстно релевантные ответы исключительно на основе предоставленного контекста из нормативной документации.
|
| 136 |
|
| 137 |
+
ПРАВИЛА АНАЛИЗА ЗАПРОСА:
|
|
|
|
| 138 |
|
| 139 |
+
1. ПРЯМЫЕ ВОПРОСЫ БЕЗ ДОКУМЕНТАЛЬНОГО КОНТЕКСТА:
|
| 140 |
+
Если пользователь задает вопрос типа "В каких случаях могут быть признаны протоколы испытаний?" без предоставления дополнительных документов, найдите соответствующую информацию в доступном контексте и предоставьте полный ответ с указанием источников.
|
| 141 |
|
| 142 |
+
2. ОПРЕДЕЛЕНИЕ ТИПА ЗАДАЧИ:
|
|
|
|
|
|
|
|
|
|
| 143 |
|
| 144 |
+
а) ПОИСК И ОТВЕТ НА ВОПРОС (ключевые слова: "в каких случаях", "когда", "кто", "что", "как", "почему"):
|
| 145 |
+
- Найдите релевантную информацию в контексте
|
| 146 |
+
- Предоставьте развернутый ответ
|
| 147 |
+
- Обязательно укажите конкретные документы и разделы
|
| 148 |
+
- Процитируйте ключевые положения
|
| 149 |
+
|
| 150 |
+
б) КРАТКОЕ САММАРИ (ключевые слова: "кратко", "суммировать", "резюме", "основные моменты"):
|
| 151 |
+
- Предоставьте структурированное резюме
|
| 152 |
+
- Выделите ключевые требования
|
| 153 |
+
- Используйте нумерованный список
|
| 154 |
+
|
| 155 |
+
в) ПОИСК ДОКУМЕНТА И ПУНКТА (ключевые слова: "найти", "где", "какой документ", "в каком разделе"):
|
| 156 |
+
- Укажите конкретный документ и структурное расположение
|
| 157 |
+
- Предоставьте точные номера разделов/пунктов
|
| 158 |
+
|
| 159 |
+
г) ПРОВЕРКА КОРРЕКТНОСТИ (ключевые слова: "правильно ли", "соответствует ли", "проверить"):
|
| 160 |
+
- Четко укажите: "СООТВЕТСТВУЕТ" или "НЕ СООТВЕТСТВУЕТ"
|
| 161 |
+
- Перечислите конкретные требования
|
| 162 |
+
|
| 163 |
+
д) ПЛАН ДЕЙСТВИЙ (ключевые слова: "план", "алгоритм", "пошагово"):
|
| 164 |
+
- Создайте пронумерованный план
|
| 165 |
+
- Укажите ссылки на соответствующие пункты НД
|
| 166 |
+
|
| 167 |
+
ПРАВИЛА ФОРМИРОВАНИЯ ОТВЕТОВ:
|
| 168 |
+
|
| 169 |
+
Работай исключительно с информацией из предоставленного контекста. Запрещено использовать:
|
| 170 |
+
- Общие знания
|
| 171 |
+
- Информацию из интернета
|
| 172 |
+
- Данные из предыдущих диалогов
|
| 173 |
+
- Собственные предположения
|
| 174 |
+
|
| 175 |
+
1. СТРУКТУРА ОТВЕТА:
|
| 176 |
+
- Начинайте с прямого ответа на вопрос
|
| 177 |
+
- Затем указывайте нормативные основания
|
| 178 |
+
- Завершайте ссылками на конкретные документы и разделы
|
| 179 |
+
|
| 180 |
+
2. РАБОТА С КОНТЕКСТОМ:
|
| 181 |
+
- Если информация найдена в контексте - предоставьте полный ответ
|
| 182 |
+
- Если информация не найдена: "Информация по вашему запросу не найдена в доступной нормативной документации"
|
| 183 |
+
- Не делайте предположений за пределами контекста
|
| 184 |
+
- Не используйте общие знания
|
| 185 |
+
|
| 186 |
+
3. ТЕРМИНОЛОГИЯ И ЦИТИРОВАНИЕ:
|
| 187 |
+
- Сохраняйте официальную терминологию НД
|
| 188 |
+
- Цитируйте точные формулировки ключевых требований
|
| 189 |
+
- При множественных источниках - укажите все релевантные
|
| 190 |
+
|
| 191 |
+
4. ФОРМАТИРОВАНИЕ:
|
| 192 |
+
- Для перечислений: используйте нумерованные списки
|
| 193 |
+
- Выделяйте критически важные требования
|
| 194 |
+
- Структурируйте ответ логически
|
| 195 |
+
|
| 196 |
+
# КАК РАБОТАТЬ С ЗАПРОСОМ
|
| 197 |
+
|
| 198 |
+
**Шаг 1:** Определи, что именно ищет пользователь (термин, требование, процедура, условие)
|
| 199 |
+
|
| 200 |
+
**Шаг 2:** Найди релевантную информацию в контексте
|
| 201 |
+
|
| 202 |
+
**Шаг 3:** Сформируй ответ:
|
| 203 |
+
- Если нашел: укажи документ и пункт, процитируй нужную часть
|
| 204 |
+
- Если не нашел: четко сообщи об отсутствии информации
|
| 205 |
+
|
| 206 |
+
**Шаг 4:** При наличии нескольких источников:
|
| 207 |
+
- Представь их последовательно с указанием источника каждого
|
| 208 |
+
- Если источников много (>4) — сначала дай их список, потом цитаты
|
| 209 |
|
| 210 |
КОНТЕКСТ:
|
| 211 |
{context}
|
| 212 |
|
| 213 |
ВОПРОС: {question}
|
| 214 |
+
"""
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 215 |
|
| 216 |
response = query_engine.query(prompt)
|
| 217 |
sources = format_sources(reranked)
|
|
|
|
| 223 |
import traceback
|
| 224 |
log_message(traceback.format_exc())
|
| 225 |
return f"Ошибка: {e}", ""
|
| 226 |
+
|
| 227 |
+
def rerank_nodes(query, nodes, reranker, top_k=30, min_score=0.3):
|
| 228 |
+
"""Rerank nodes with diversity - MORE LENIENT"""
|
| 229 |
if not nodes:
|
| 230 |
return []
|
| 231 |
|
| 232 |
+
# Score all nodes
|
| 233 |
pairs = [[query, n.text] for n in nodes]
|
| 234 |
scores = reranker.predict(pairs)
|
| 235 |
|
| 236 |
+
# Sort by score
|
| 237 |
scored = sorted(zip(nodes, scores), key=lambda x: x[1], reverse=True)
|
| 238 |
|
| 239 |
+
# More lenient threshold
|
|
|
|
|
|
|
| 240 |
filtered = [(n, s) for n, s in scored if s >= min_score]
|
| 241 |
|
| 242 |
if not filtered:
|
| 243 |
+
# Fallback: take top 50% if nothing passes threshold
|
| 244 |
+
cutoff = max(scores) * 0.5
|
| 245 |
filtered = [(n, s) for n, s in scored if s >= cutoff][:top_k]
|
| 246 |
|
| 247 |
+
# Log top scores for debugging
|
| 248 |
+
log_message(f"Top 5 reranking scores: {[f'{s:.3f}' for _, s in scored[:5]]}")
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 249 |
|
| 250 |
+
# Diversity selection - but prioritize tables if query mentions them
|
| 251 |
selected = []
|
| 252 |
+
seen_docs = set()
|
| 253 |
+
table_nodes = []
|
| 254 |
+
other_nodes = []
|
| 255 |
+
|
| 256 |
+
for node, score in filtered:
|
| 257 |
+
if node.metadata.get('type') == 'table':
|
| 258 |
+
table_nodes.append((node, score))
|
| 259 |
+
else:
|
| 260 |
+
other_nodes.append((node, score))
|
| 261 |
|
| 262 |
+
# If query mentions "таблица", prioritize table nodes
|
| 263 |
+
if 'таблиц' in query.lower():
|
| 264 |
+
combined = table_nodes + other_nodes
|
| 265 |
+
else:
|
| 266 |
+
combined = filtered
|
| 267 |
|
| 268 |
+
for node, score in combined[:top_k]:
|
| 269 |
+
if len(selected) >= top_k:
|
| 270 |
+
break
|
| 271 |
+
selected.append(node)
|
| 272 |
+
seen_docs.add(node.metadata.get('document_id', 'unknown'))
|
| 273 |
|
| 274 |
+
log_message(f"Reranked: {len(filtered)} → {len(selected)} (from {len(seen_docs)} docs)")
|
| 275 |
|
| 276 |
+
return selected
|