Spaces:
Sleeping
Sleeping
Commit
·
f85ad1c
1
Parent(s):
822ef8c
new way of chunking
Browse files- config.py +1 -1
- documents_prep.py +127 -68
- table_prep.py +106 -60
config.py
CHANGED
|
@@ -50,7 +50,7 @@ AVAILABLE_MODELS = {
|
|
| 50 |
|
| 51 |
DEFAULT_MODEL = "Gemini 2.5 Flash"
|
| 52 |
|
| 53 |
-
CHUNK_SIZE =
|
| 54 |
CHUNK_OVERLAP = 256
|
| 55 |
|
| 56 |
CUSTOM_PROMPT = """
|
|
|
|
| 50 |
|
| 51 |
DEFAULT_MODEL = "Gemini 2.5 Flash"
|
| 52 |
|
| 53 |
+
CHUNK_SIZE = 2000
|
| 54 |
CHUNK_OVERLAP = 256
|
| 55 |
|
| 56 |
CUSTOM_PROMPT = """
|
documents_prep.py
CHANGED
|
@@ -14,147 +14,206 @@ def chunk_document(doc, chunk_size=None, chunk_overlap=None):
|
|
| 14 |
chunk_size = CHUNK_SIZE
|
| 15 |
if chunk_overlap is None:
|
| 16 |
chunk_overlap = CHUNK_OVERLAP
|
| 17 |
-
text_splitter = SentenceSplitter(
|
| 18 |
-
chunk_size=chunk_size,
|
| 19 |
-
chunk_overlap=chunk_overlap,
|
| 20 |
-
separator=" "
|
| 21 |
-
)
|
| 22 |
|
| 23 |
-
|
| 24 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 25 |
chunked_docs = []
|
| 26 |
-
for i, chunk_text in enumerate(
|
| 27 |
chunk_metadata = doc.metadata.copy()
|
| 28 |
chunk_metadata.update({
|
| 29 |
"chunk_id": i,
|
| 30 |
-
"total_chunks": len(
|
| 31 |
"chunk_size": len(chunk_text),
|
| 32 |
-
"
|
| 33 |
})
|
| 34 |
-
|
| 35 |
-
chunked_doc = Document(
|
| 36 |
-
text=chunk_text,
|
| 37 |
-
metadata=chunk_metadata
|
| 38 |
-
)
|
| 39 |
-
chunked_docs.append(chunked_doc)
|
| 40 |
|
| 41 |
return chunked_docs
|
| 42 |
|
| 43 |
def process_documents_with_chunking(documents):
|
|
|
|
|
|
|
|
|
|
|
|
|
| 44 |
all_chunked_docs = []
|
| 45 |
chunk_info = []
|
| 46 |
-
table_count = 0
|
| 47 |
-
table_chunks_count = 0
|
| 48 |
-
image_count = 0
|
| 49 |
-
image_chunks_count = 0
|
| 50 |
-
text_chunks_count = 0
|
| 51 |
|
| 52 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 53 |
doc_type = doc.metadata.get('type', 'text')
|
| 54 |
is_already_chunked = doc.metadata.get('is_chunked', False)
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 55 |
|
| 56 |
if doc_type == 'table':
|
| 57 |
if is_already_chunked:
|
| 58 |
-
|
|
|
|
| 59 |
all_chunked_docs.append(doc)
|
| 60 |
-
|
| 61 |
-
|
| 62 |
-
'section_id': doc.metadata.get('section_id', 'unknown'),
|
| 63 |
-
'chunk_id': doc.metadata.get('chunk_id', 0),
|
| 64 |
-
'total_chunks': doc.metadata.get('total_chunks', 1),
|
| 65 |
-
'chunk_size': len(doc.text),
|
| 66 |
-
'chunk_preview': doc.text[:200] + "..." if len(doc.text) > 200 else doc.text,
|
| 67 |
-
'type': 'table',
|
| 68 |
-
'table_number': doc.metadata.get('table_number', 'unknown')
|
| 69 |
-
})
|
| 70 |
else:
|
| 71 |
-
|
|
|
|
| 72 |
all_chunked_docs.append(doc)
|
| 73 |
-
|
| 74 |
-
|
| 75 |
-
|
| 76 |
-
|
| 77 |
-
|
| 78 |
-
|
| 79 |
-
|
| 80 |
-
|
| 81 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 82 |
|
| 83 |
elif doc_type == 'image':
|
| 84 |
-
image_count += 1
|
| 85 |
-
doc_size = len(doc.text)
|
| 86 |
if doc_size > CHUNK_SIZE:
|
| 87 |
-
log_message(f"📷
|
| 88 |
-
f"Размер: {doc_size} > {CHUNK_SIZE}")
|
| 89 |
chunked_docs = chunk_document(doc)
|
| 90 |
-
|
| 91 |
all_chunked_docs.extend(chunked_docs)
|
| 92 |
-
log_message(f" ✂️ Разделено на {len(chunked_docs)} чанков")
|
| 93 |
|
| 94 |
-
for
|
| 95 |
chunk_info.append({
|
| 96 |
'document_id': chunk_doc.metadata.get('document_id', 'unknown'),
|
| 97 |
'section_id': chunk_doc.metadata.get('section_id', 'unknown'),
|
| 98 |
-
'chunk_id':
|
|
|
|
| 99 |
'chunk_size': len(chunk_doc.text),
|
| 100 |
'chunk_preview': chunk_doc.text[:200] + "..." if len(chunk_doc.text) > 200 else chunk_doc.text,
|
| 101 |
'type': 'image',
|
| 102 |
-
'image_number': chunk_doc.metadata.get('image_number', 'unknown')
|
|
|
|
| 103 |
})
|
| 104 |
else:
|
|
|
|
| 105 |
all_chunked_docs.append(doc)
|
|
|
|
|
|
|
| 106 |
chunk_info.append({
|
| 107 |
'document_id': doc.metadata.get('document_id', 'unknown'),
|
| 108 |
'section_id': doc.metadata.get('section_id', 'unknown'),
|
| 109 |
'chunk_id': 0,
|
|
|
|
| 110 |
'chunk_size': doc_size,
|
| 111 |
'chunk_preview': doc.text[:200] + "..." if len(doc.text) > 200 else doc.text,
|
| 112 |
'type': 'image',
|
| 113 |
-
'image_number': doc.metadata.get('image_number', 'unknown')
|
|
|
|
| 114 |
})
|
| 115 |
|
| 116 |
-
else:
|
| 117 |
-
doc_size = len(doc.text)
|
| 118 |
if doc_size > CHUNK_SIZE:
|
| 119 |
-
log_message(f"📝
|
|
|
|
|
|
|
| 120 |
f"Размер: {doc_size} > {CHUNK_SIZE}")
|
|
|
|
| 121 |
chunked_docs = chunk_document(doc)
|
| 122 |
-
|
| 123 |
all_chunked_docs.extend(chunked_docs)
|
| 124 |
-
log_message(f" ✂️ Разделен на {len(chunked_docs)} чанков")
|
| 125 |
|
| 126 |
-
for
|
| 127 |
chunk_info.append({
|
| 128 |
'document_id': chunk_doc.metadata.get('document_id', 'unknown'),
|
| 129 |
'section_id': chunk_doc.metadata.get('section_id', 'unknown'),
|
| 130 |
-
'chunk_id':
|
|
|
|
| 131 |
'chunk_size': len(chunk_doc.text),
|
| 132 |
'chunk_preview': chunk_doc.text[:200] + "..." if len(chunk_doc.text) > 200 else chunk_doc.text,
|
| 133 |
-
'type': 'text'
|
|
|
|
| 134 |
})
|
| 135 |
else:
|
|
|
|
| 136 |
all_chunked_docs.append(doc)
|
|
|
|
|
|
|
| 137 |
chunk_info.append({
|
| 138 |
'document_id': doc.metadata.get('document_id', 'unknown'),
|
| 139 |
'section_id': doc.metadata.get('section_id', 'unknown'),
|
| 140 |
'chunk_id': 0,
|
|
|
|
| 141 |
'chunk_size': doc_size,
|
| 142 |
'chunk_preview': doc.text[:200] + "..." if len(doc.text) > 200 else doc.text,
|
| 143 |
-
'type': 'text'
|
|
|
|
| 144 |
})
|
| 145 |
|
| 146 |
log_message(f"\n{'='*60}")
|
| 147 |
-
log_message(f"
|
| 148 |
-
log_message(f"
|
| 149 |
-
log_message(f"
|
| 150 |
-
log_message(f"
|
| 151 |
-
log_message(f"
|
| 152 |
-
log_message(f"
|
| 153 |
-
log_message(f"
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 154 |
log_message(f"{'='*60}\n")
|
| 155 |
|
| 156 |
return all_chunked_docs, chunk_info
|
| 157 |
|
|
|
|
| 158 |
def extract_text_from_json(data, document_id, document_name):
|
| 159 |
documents = []
|
| 160 |
|
|
|
|
| 14 |
chunk_size = CHUNK_SIZE
|
| 15 |
if chunk_overlap is None:
|
| 16 |
chunk_overlap = CHUNK_OVERLAP
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 17 |
|
| 18 |
+
text = doc.text
|
| 19 |
|
| 20 |
+
# Try to split by double newlines (paragraphs) first
|
| 21 |
+
paragraphs = text.split('\n\n')
|
| 22 |
+
|
| 23 |
+
chunks = []
|
| 24 |
+
current_chunk = ""
|
| 25 |
+
|
| 26 |
+
for para in paragraphs:
|
| 27 |
+
para = para.strip()
|
| 28 |
+
if not para:
|
| 29 |
+
continue
|
| 30 |
+
|
| 31 |
+
# If adding this paragraph exceeds limit, save current chunk
|
| 32 |
+
if len(current_chunk) + len(para) + 2 > chunk_size and current_chunk:
|
| 33 |
+
chunks.append(current_chunk.strip())
|
| 34 |
+
# Add overlap from end of previous chunk
|
| 35 |
+
overlap_text = current_chunk[-chunk_overlap:] if len(current_chunk) > chunk_overlap else current_chunk
|
| 36 |
+
current_chunk = overlap_text + "\n\n" + para
|
| 37 |
+
else:
|
| 38 |
+
if current_chunk:
|
| 39 |
+
current_chunk += "\n\n" + para
|
| 40 |
+
else:
|
| 41 |
+
current_chunk = para
|
| 42 |
+
|
| 43 |
+
# Add last chunk
|
| 44 |
+
if current_chunk:
|
| 45 |
+
chunks.append(current_chunk.strip())
|
| 46 |
+
|
| 47 |
+
# If single paragraph is too large, fall back to sentence splitting
|
| 48 |
+
final_chunks = []
|
| 49 |
+
for chunk_text in chunks:
|
| 50 |
+
if len(chunk_text) > chunk_size:
|
| 51 |
+
splitter = SentenceSplitter(chunk_size=chunk_size, chunk_overlap=chunk_overlap)
|
| 52 |
+
final_chunks.extend(splitter.split_text(chunk_text))
|
| 53 |
+
else:
|
| 54 |
+
final_chunks.append(chunk_text)
|
| 55 |
+
|
| 56 |
+
log_message(f" ✂️ Текст разбит на {len(final_chunks)} семантических чанков")
|
| 57 |
+
|
| 58 |
+
# Create documents
|
| 59 |
chunked_docs = []
|
| 60 |
+
for i, chunk_text in enumerate(final_chunks):
|
| 61 |
chunk_metadata = doc.metadata.copy()
|
| 62 |
chunk_metadata.update({
|
| 63 |
"chunk_id": i,
|
| 64 |
+
"total_chunks": len(final_chunks),
|
| 65 |
"chunk_size": len(chunk_text),
|
| 66 |
+
"is_chunked": True
|
| 67 |
})
|
| 68 |
+
chunked_docs.append(Document(text=chunk_text, metadata=chunk_metadata))
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 69 |
|
| 70 |
return chunked_docs
|
| 71 |
|
| 72 |
def process_documents_with_chunking(documents):
|
| 73 |
+
log_message("\n" + "="*60)
|
| 74 |
+
log_message("🔄 НАЧАЛО ПРОЦЕССА ЧАНКИНГА")
|
| 75 |
+
log_message("="*60)
|
| 76 |
+
|
| 77 |
all_chunked_docs = []
|
| 78 |
chunk_info = []
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 79 |
|
| 80 |
+
# Counters
|
| 81 |
+
table_whole_count = 0 # Целые таблицы (не нуждаются в чанкинге)
|
| 82 |
+
table_chunked_count = 0 # Таблицы, которые УЖЕ разбиты
|
| 83 |
+
image_whole_count = 0 # Целые изображения
|
| 84 |
+
image_chunked_count = 0 # Изображения, разбитые на чанки
|
| 85 |
+
text_whole_count = 0 # Целые текстовые документы
|
| 86 |
+
text_chunked_count = 0 # Текстовые документы, разбитые на чанки
|
| 87 |
+
|
| 88 |
+
for idx, doc in enumerate(documents):
|
| 89 |
doc_type = doc.metadata.get('type', 'text')
|
| 90 |
is_already_chunked = doc.metadata.get('is_chunked', False)
|
| 91 |
+
doc_size = len(doc.text)
|
| 92 |
+
|
| 93 |
+
log_message(f"\n📄 Документ {idx+1}/{len(documents)} | "
|
| 94 |
+
f"Тип: {doc_type} | "
|
| 95 |
+
f"Размер: {doc_size} | "
|
| 96 |
+
f"Уже разбит: {is_already_chunked}")
|
| 97 |
|
| 98 |
if doc_type == 'table':
|
| 99 |
if is_already_chunked:
|
| 100 |
+
# Таблица уже разбита на чанки в table_prep.py
|
| 101 |
+
table_chunked_count += 1
|
| 102 |
all_chunked_docs.append(doc)
|
| 103 |
+
log_message(f" ✓ Таблица (чанк {doc.metadata.get('chunk_id', 0) + 1}/"
|
| 104 |
+
f"{doc.metadata.get('total_chunks', 1)}) добавлена без изменений")
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 105 |
else:
|
| 106 |
+
# Целая таблица
|
| 107 |
+
table_whole_count += 1
|
| 108 |
all_chunked_docs.append(doc)
|
| 109 |
+
log_message(f" ✓ Целая таблица добавлена | "
|
| 110 |
+
f"Номер: {doc.metadata.get('table_number', 'unknown')}")
|
| 111 |
+
|
| 112 |
+
chunk_info.append({
|
| 113 |
+
'document_id': doc.metadata.get('document_id', 'unknown'),
|
| 114 |
+
'section_id': doc.metadata.get('section_id', 'unknown'),
|
| 115 |
+
'chunk_id': doc.metadata.get('chunk_id', 0),
|
| 116 |
+
'total_chunks': doc.metadata.get('total_chunks', 1),
|
| 117 |
+
'chunk_size': doc_size,
|
| 118 |
+
'chunk_preview': doc.text[:200] + "..." if len(doc.text) > 200 else doc.text,
|
| 119 |
+
'type': 'table',
|
| 120 |
+
'table_number': doc.metadata.get('table_number', 'unknown'),
|
| 121 |
+
'is_chunked': is_already_chunked
|
| 122 |
+
})
|
| 123 |
|
| 124 |
elif doc_type == 'image':
|
|
|
|
|
|
|
| 125 |
if doc_size > CHUNK_SIZE:
|
| 126 |
+
log_message(f" 📷 Изображение требует чанкинга | Размер: {doc_size} > {CHUNK_SIZE}")
|
|
|
|
| 127 |
chunked_docs = chunk_document(doc)
|
| 128 |
+
image_chunked_count += len(chunked_docs)
|
| 129 |
all_chunked_docs.extend(chunked_docs)
|
|
|
|
| 130 |
|
| 131 |
+
for chunk_doc in chunked_docs:
|
| 132 |
chunk_info.append({
|
| 133 |
'document_id': chunk_doc.metadata.get('document_id', 'unknown'),
|
| 134 |
'section_id': chunk_doc.metadata.get('section_id', 'unknown'),
|
| 135 |
+
'chunk_id': chunk_doc.metadata.get('chunk_id', 0),
|
| 136 |
+
'total_chunks': chunk_doc.metadata.get('total_chunks', 1),
|
| 137 |
'chunk_size': len(chunk_doc.text),
|
| 138 |
'chunk_preview': chunk_doc.text[:200] + "..." if len(chunk_doc.text) > 200 else chunk_doc.text,
|
| 139 |
'type': 'image',
|
| 140 |
+
'image_number': chunk_doc.metadata.get('image_number', 'unknown'),
|
| 141 |
+
'is_chunked': True
|
| 142 |
})
|
| 143 |
else:
|
| 144 |
+
image_whole_count += 1
|
| 145 |
all_chunked_docs.append(doc)
|
| 146 |
+
log_message(f" ✓ Целое изображение добавлено | Размер: {doc_size}")
|
| 147 |
+
|
| 148 |
chunk_info.append({
|
| 149 |
'document_id': doc.metadata.get('document_id', 'unknown'),
|
| 150 |
'section_id': doc.metadata.get('section_id', 'unknown'),
|
| 151 |
'chunk_id': 0,
|
| 152 |
+
'total_chunks': 1,
|
| 153 |
'chunk_size': doc_size,
|
| 154 |
'chunk_preview': doc.text[:200] + "..." if len(doc.text) > 200 else doc.text,
|
| 155 |
'type': 'image',
|
| 156 |
+
'image_number': doc.metadata.get('image_number', 'unknown'),
|
| 157 |
+
'is_chunked': False
|
| 158 |
})
|
| 159 |
|
| 160 |
+
else: # text
|
|
|
|
| 161 |
if doc_size > CHUNK_SIZE:
|
| 162 |
+
log_message(f" 📝 Текст требует чанкинга | "
|
| 163 |
+
f"Документ: {doc.metadata.get('document_id', 'unknown')} | "
|
| 164 |
+
f"Раздел: {doc.metadata.get('section_id', 'unknown')} | "
|
| 165 |
f"Размер: {doc_size} > {CHUNK_SIZE}")
|
| 166 |
+
|
| 167 |
chunked_docs = chunk_document(doc)
|
| 168 |
+
text_chunked_count += len(chunked_docs)
|
| 169 |
all_chunked_docs.extend(chunked_docs)
|
|
|
|
| 170 |
|
| 171 |
+
for chunk_doc in chunked_docs:
|
| 172 |
chunk_info.append({
|
| 173 |
'document_id': chunk_doc.metadata.get('document_id', 'unknown'),
|
| 174 |
'section_id': chunk_doc.metadata.get('section_id', 'unknown'),
|
| 175 |
+
'chunk_id': chunk_doc.metadata.get('chunk_id', 0),
|
| 176 |
+
'total_chunks': chunk_doc.metadata.get('total_chunks', 1),
|
| 177 |
'chunk_size': len(chunk_doc.text),
|
| 178 |
'chunk_preview': chunk_doc.text[:200] + "..." if len(chunk_doc.text) > 200 else chunk_doc.text,
|
| 179 |
+
'type': 'text',
|
| 180 |
+
'is_chunked': True
|
| 181 |
})
|
| 182 |
else:
|
| 183 |
+
text_whole_count += 1
|
| 184 |
all_chunked_docs.append(doc)
|
| 185 |
+
log_message(f" ✓ Целый текстовый документ добавлен | Размер: {doc_size}")
|
| 186 |
+
|
| 187 |
chunk_info.append({
|
| 188 |
'document_id': doc.metadata.get('document_id', 'unknown'),
|
| 189 |
'section_id': doc.metadata.get('section_id', 'unknown'),
|
| 190 |
'chunk_id': 0,
|
| 191 |
+
'total_chunks': 1,
|
| 192 |
'chunk_size': doc_size,
|
| 193 |
'chunk_preview': doc.text[:200] + "..." if len(doc.text) > 200 else doc.text,
|
| 194 |
+
'type': 'text',
|
| 195 |
+
'is_chunked': False
|
| 196 |
})
|
| 197 |
|
| 198 |
log_message(f"\n{'='*60}")
|
| 199 |
+
log_message(f"📊 ИТОГОВАЯ СТАТИСТИКА ЧАНКИНГА:")
|
| 200 |
+
log_message(f"{'='*60}")
|
| 201 |
+
log_message(f" ТАБЛИЦЫ:")
|
| 202 |
+
log_message(f" • Целые (не нуждались в чанкинге): {table_whole_count}")
|
| 203 |
+
log_message(f" • Чанки (разбиты в table_prep.py): {table_chunked_count}")
|
| 204 |
+
log_message(f" ИЗОБРАЖЕНИЯ:")
|
| 205 |
+
log_message(f" • Целые: {image_whole_count}")
|
| 206 |
+
log_message(f" • Чанки: {image_chunked_count}")
|
| 207 |
+
log_message(f" ТЕКСТ:")
|
| 208 |
+
log_message(f" • Целые документы: {text_whole_count}")
|
| 209 |
+
log_message(f" • Чанки: {text_chunked_count}")
|
| 210 |
+
log_message(f" {'─'*58}")
|
| 211 |
+
log_message(f" ВСЕГО ДОКУМЕНТОВ В ИНДЕКСЕ: {len(all_chunked_docs)}")
|
| 212 |
log_message(f"{'='*60}\n")
|
| 213 |
|
| 214 |
return all_chunked_docs, chunk_info
|
| 215 |
|
| 216 |
+
|
| 217 |
def extract_text_from_json(data, document_id, document_name):
|
| 218 |
documents = []
|
| 219 |
|
table_prep.py
CHANGED
|
@@ -32,39 +32,80 @@ def create_table_content(table_data):
|
|
| 32 |
from llama_index.core.text_splitter import SentenceSplitter
|
| 33 |
from config import CHUNK_SIZE, CHUNK_OVERLAP
|
| 34 |
|
| 35 |
-
def
|
| 36 |
-
|
| 37 |
-
|
| 38 |
-
|
| 39 |
-
|
| 40 |
-
|
| 41 |
-
|
| 42 |
-
|
| 43 |
-
|
| 44 |
-
|
| 45 |
-
|
| 46 |
-
|
| 47 |
-
|
| 48 |
-
|
| 49 |
-
|
| 50 |
-
|
| 51 |
-
|
| 52 |
-
|
| 53 |
-
|
| 54 |
-
|
| 55 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 56 |
"is_chunked": True
|
| 57 |
-
}
|
| 58 |
|
| 59 |
-
|
| 60 |
-
|
| 61 |
-
|
| 62 |
-
)
|
| 63 |
-
|
| 64 |
|
| 65 |
-
return
|
|
|
|
| 66 |
|
| 67 |
def table_to_document(table_data, document_id=None):
|
|
|
|
|
|
|
|
|
|
| 68 |
if not isinstance(table_data, dict):
|
| 69 |
log_message(f"⚠️ ПРОПУЩЕНА: table_data не является словарем")
|
| 70 |
return []
|
|
@@ -75,41 +116,46 @@ def table_to_document(table_data, document_id=None):
|
|
| 75 |
section = table_data.get('section', 'Неизвестно')
|
| 76 |
|
| 77 |
table_rows = table_data.get('data', [])
|
| 78 |
-
if not table_rows
|
| 79 |
-
log_message(f"⚠️ ПРОПУЩЕНА: Таблица {table_num} из '{doc_id}' - нет данных
|
| 80 |
return []
|
| 81 |
|
| 82 |
-
|
| 83 |
-
|
| 84 |
-
|
| 85 |
-
|
| 86 |
-
|
| 87 |
-
|
| 88 |
-
|
| 89 |
-
|
| 90 |
-
|
| 91 |
-
|
| 92 |
-
|
| 93 |
-
|
| 94 |
-
|
| 95 |
-
|
| 96 |
-
|
| 97 |
-
|
| 98 |
-
|
| 99 |
-
|
| 100 |
-
|
| 101 |
-
|
| 102 |
-
|
| 103 |
-
|
| 104 |
-
|
| 105 |
-
|
| 106 |
-
|
| 107 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
| 108 |
else:
|
| 109 |
-
log_message(f"
|
| 110 |
-
|
| 111 |
-
|
| 112 |
-
|
|
|
|
| 113 |
|
| 114 |
def load_table_data(repo_id, hf_token, table_data_dir):
|
| 115 |
log_message("=" * 60)
|
|
|
|
| 32 |
from llama_index.core.text_splitter import SentenceSplitter
|
| 33 |
from config import CHUNK_SIZE, CHUNK_OVERLAP
|
| 34 |
|
| 35 |
+
def create_table_chunks_with_headers(table_data, rows_per_chunk=10):
|
| 36 |
+
"""
|
| 37 |
+
Intelligently chunk tables by preserving headers and grouping rows
|
| 38 |
+
"""
|
| 39 |
+
doc_id = table_data.get('document_id') or table_data.get('document', 'Неизвестно')
|
| 40 |
+
table_num = table_data.get('table_number', 'Неизвестно')
|
| 41 |
+
table_title = table_data.get('table_title', 'Неизвестно')
|
| 42 |
+
section = table_data.get('section', 'Неизвестно')
|
| 43 |
+
headers = table_data.get('headers', [])
|
| 44 |
+
table_rows = table_data.get('data', [])
|
| 45 |
+
|
| 46 |
+
if not table_rows:
|
| 47 |
+
return []
|
| 48 |
+
|
| 49 |
+
# Create header string that will be included in EVERY chunk
|
| 50 |
+
header_context = f"Таблица {table_num}: {table_title}\n"
|
| 51 |
+
header_context += f"Документ: {doc_id}\n"
|
| 52 |
+
header_context += f"Раздел: {section}\n"
|
| 53 |
+
if headers:
|
| 54 |
+
header_context += f"Заголовки: {' | '.join(headers)}\n"
|
| 55 |
+
header_context += f"Всего строк в таблице: {len(table_rows)}\n\n"
|
| 56 |
+
|
| 57 |
+
# Calculate optimal rows per chunk based on content size
|
| 58 |
+
avg_row_size = sum(len(str(row)) for row in table_rows[:5]) / min(5, len(table_rows))
|
| 59 |
+
max_chunk_size = CHUNK_SIZE - len(header_context) - 500 # Safety margin
|
| 60 |
+
optimal_rows = max(5, int(max_chunk_size / avg_row_size))
|
| 61 |
+
|
| 62 |
+
log_message(f" 📐 Средний размер строки: {avg_row_size:.0f} символов")
|
| 63 |
+
log_message(f" 📊 Оптимальное кол-во строк на чанк: {optimal_rows}")
|
| 64 |
+
|
| 65 |
+
chunks = []
|
| 66 |
+
total_rows = len(table_rows)
|
| 67 |
+
|
| 68 |
+
for i in range(0, total_rows, optimal_rows):
|
| 69 |
+
chunk_rows = table_rows[i:i + optimal_rows]
|
| 70 |
+
|
| 71 |
+
# Build chunk content
|
| 72 |
+
chunk_content = header_context
|
| 73 |
+
chunk_content += f"[Строки {i+1}-{min(i+optimal_rows, total_rows)} из {total_rows}]\n"
|
| 74 |
+
chunk_content += "Данные:\n"
|
| 75 |
+
|
| 76 |
+
for row_idx, row in enumerate(chunk_rows, start=i+1):
|
| 77 |
+
if isinstance(row, dict):
|
| 78 |
+
row_text = " | ".join([f"{k}: {v}" for k, v in row.items() if v])
|
| 79 |
+
chunk_content += f"Строка {row_idx}: {row_text}\n"
|
| 80 |
+
|
| 81 |
+
chunk_metadata = {
|
| 82 |
+
"type": "table",
|
| 83 |
+
"table_number": table_num,
|
| 84 |
+
"table_title": table_title,
|
| 85 |
+
"document_id": doc_id,
|
| 86 |
+
"section": section,
|
| 87 |
+
"section_id": section,
|
| 88 |
+
"headers": headers,
|
| 89 |
+
"chunk_id": i // optimal_rows,
|
| 90 |
+
"total_chunks": (total_rows + optimal_rows - 1) // optimal_rows,
|
| 91 |
+
"row_range": f"{i+1}-{min(i+optimal_rows, total_rows)}",
|
| 92 |
+
"total_table_rows": total_rows,
|
| 93 |
"is_chunked": True
|
| 94 |
+
}
|
| 95 |
|
| 96 |
+
doc = Document(text=chunk_content, metadata=chunk_metadata)
|
| 97 |
+
chunks.append(doc)
|
| 98 |
+
|
| 99 |
+
log_message(f" Чанк {len(chunks)}: строки {i+1}-{min(i+optimal_rows, total_rows)} | "
|
| 100 |
+
f"{len(chunk_content)} символов")
|
| 101 |
|
| 102 |
+
return chunks
|
| 103 |
+
|
| 104 |
|
| 105 |
def table_to_document(table_data, document_id=None):
|
| 106 |
+
"""
|
| 107 |
+
Convert table to Document(s) with intelligent chunking
|
| 108 |
+
"""
|
| 109 |
if not isinstance(table_data, dict):
|
| 110 |
log_message(f"⚠️ ПРОПУЩЕНА: table_data не является словарем")
|
| 111 |
return []
|
|
|
|
| 116 |
section = table_data.get('section', 'Неизвестно')
|
| 117 |
|
| 118 |
table_rows = table_data.get('data', [])
|
| 119 |
+
if not table_rows:
|
| 120 |
+
log_message(f"⚠️ ПРОПУЩЕНА: Таблица {table_num} из '{doc_id}' - нет данных")
|
| 121 |
return []
|
| 122 |
|
| 123 |
+
log_message(f"\n📊 Обработка таблицы {table_num} из документа '{doc_id}'")
|
| 124 |
+
log_message(f" Название: {table_title}")
|
| 125 |
+
log_message(f" Раздел: {section}")
|
| 126 |
+
log_message(f" Строк данных: {len(table_rows)}")
|
| 127 |
+
|
| 128 |
+
# Estimate if table needs chunking
|
| 129 |
+
sample_content = create_table_content(table_data)
|
| 130 |
+
estimated_size = len(sample_content)
|
| 131 |
+
|
| 132 |
+
log_message(f" Оценочный размер: {estimated_size} символов")
|
| 133 |
+
|
| 134 |
+
# Threshold: if table is small enough, keep it whole
|
| 135 |
+
if estimated_size <= CHUNK_SIZE * 0.8: # 80% of limit for safety
|
| 136 |
+
log_message(f" ✅ Таблица достаточно мала, хранится целиком")
|
| 137 |
+
doc = Document(
|
| 138 |
+
text=sample_content,
|
| 139 |
+
metadata={
|
| 140 |
+
"type": "table",
|
| 141 |
+
"table_number": table_num,
|
| 142 |
+
"table_title": table_title,
|
| 143 |
+
"document_id": doc_id,
|
| 144 |
+
"section": section,
|
| 145 |
+
"section_id": section,
|
| 146 |
+
"headers": table_data.get('headers', []),
|
| 147 |
+
"total_rows": len(table_rows),
|
| 148 |
+
"content_size": estimated_size,
|
| 149 |
+
"is_chunked": False
|
| 150 |
+
}
|
| 151 |
+
)
|
| 152 |
+
return [doc]
|
| 153 |
else:
|
| 154 |
+
log_message(f" ⚠️ Таблица слишком большая ({estimated_size} > {CHUNK_SIZE})")
|
| 155 |
+
log_message(f" 🔄 Применяется умный чанкинг с сохранением заголовков...")
|
| 156 |
+
chunks = create_table_chunks_with_headers(table_data)
|
| 157 |
+
log_message(f" ✅ Таблица разбита на {len(chunks)} чанков с сохранением структуры")
|
| 158 |
+
return chunks
|
| 159 |
|
| 160 |
def load_table_data(repo_id, hf_token, table_data_dir):
|
| 161 |
log_message("=" * 60)
|