pengjunjie/ragflow
1
0
Fork 1
mirror of https://github.com/infiniflow/ragflow.git synced 2026-01-02 10:42:36 +08:00
Code Issues Packages Projects Releases Wiki Activity

Format file format from Windows/dos to Unix (#1949)

Browse Source 
### What problem does this PR solve?

Related source file is in Windows/DOS format, they are format to Unix
format.

### Type of change

- [x] Refactoring

Signed-off-by: Jin Hai <haijin.chn@gmail.com>
This commit is contained in:
Jin Hai
2024-08-15 09:17:36 +08:00
committed by GitHub
parent 1328d715db
commit 6b3a40be5c
108 changed files with 36399 additions and 36399 deletions
Download Patch File Download Diff File Expand all files Collapse all files

318
rag/app/book.py
View File

@ -1,159 +1,159 @@
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
#
import copy
from tika import parser
import re
from io import BytesIO
from rag.nlp import bullets_category, is_english, tokenize, remove_contents_table, \
hierarchical_merge, make_colon_as_title, naive_merge, random_choices, tokenize_table, add_positions, \
tokenize_chunks, find_codec
from rag.nlp import rag_tokenizer
from deepdoc.parser import PdfParser, DocxParser, PlainParser, HtmlParser
class Pdf(PdfParser):
def __call__(self, filename, binary=None, from_page=0,
to_page=100000, zoomin=3, callback=None):
callback(msg="OCR is running...")
self.__images__(
filename if not binary else binary,
zoomin,
from_page,
to_page,
callback)
callback(msg="OCR finished")
from timeit import default_timer as timer
start = timer()
self._layouts_rec(zoomin)
callback(0.67, "Layout analysis finished")
print("layouts:", timer() - start)
self._table_transformer_job(zoomin)
callback(0.68, "Table analysis finished")
self._text_merge()
tbls = self._extract_table_figure(True, zoomin, True, True)
self._naive_vertical_merge()
self._filter_forpages()
self._merge_with_same_bullet()
callback(0.75, "Text merging finished.")
callback(0.8, "Text extraction finished")
return [(b["text"] + self._line_tag(b, zoomin), b.get("layoutno", ""))
for b in self.boxes], tbls
def chunk(filename, binary=None, from_page=0, to_page=100000,
lang="Chinese", callback=None, **kwargs):
"""
Supported file formats are docx, pdf, txt.
Since a book is long and not all the parts are useful, if it's a PDF,
please setup the page ranges for every book in order eliminate negative effects and save elapsed computing time.
"""
doc = {
"docnm_kwd": filename,
"title_tks": rag_tokenizer.tokenize(re.sub(r"\.[a-zA-Z]+$", "", filename))
}
doc["title_sm_tks"] = rag_tokenizer.fine_grained_tokenize(doc["title_tks"])
pdf_parser = None
sections, tbls = [], []
if re.search(r"\.docx$", filename, re.IGNORECASE):
callback(0.1, "Start to parse.")
doc_parser = DocxParser()
# TODO: table of contents need to be removed
sections, tbls = doc_parser(
binary if binary else filename, from_page=from_page, to_page=to_page)
remove_contents_table(sections, eng=is_english(
random_choices([t for t, _ in sections], k=200)))
tbls = [((None, lns), None) for lns in tbls]
callback(0.8, "Finish parsing.")
elif re.search(r"\.pdf$", filename, re.IGNORECASE):
pdf_parser = Pdf() if kwargs.get(
"parser_config", {}).get(
"layout_recognize", True) else PlainParser()
sections, tbls = pdf_parser(filename if not binary else binary,
from_page=from_page, to_page=to_page, callback=callback)
elif re.search(r"\.txt$", filename, re.IGNORECASE):
callback(0.1, "Start to parse.")
txt = ""
if binary:
encoding = find_codec(binary)
txt = binary.decode(encoding, errors="ignore")
else:
with open(filename, "r") as f:
while True:
l = f.readline()
if not l:
break
txt += l
sections = txt.split("\n")
sections = [(l, "") for l in sections if l]
remove_contents_table(sections, eng=is_english(
random_choices([t for t, _ in sections], k=200)))
callback(0.8, "Finish parsing.")
elif re.search(r"\.(htm|html)$", filename, re.IGNORECASE):
callback(0.1, "Start to parse.")
sections = HtmlParser()(filename, binary)
sections = [(l, "") for l in sections if l]
remove_contents_table(sections, eng=is_english(
random_choices([t for t, _ in sections], k=200)))
callback(0.8, "Finish parsing.")
elif re.search(r"\.doc$", filename, re.IGNORECASE):
callback(0.1, "Start to parse.")
binary = BytesIO(binary)
doc_parsed = parser.from_buffer(binary)
sections = doc_parsed['content'].split('\n')
sections = [(l, "") for l in sections if l]
remove_contents_table(sections, eng=is_english(
random_choices([t for t, _ in sections], k=200)))
callback(0.8, "Finish parsing.")
else:
raise NotImplementedError(
"file type not supported yet(doc, docx, pdf, txt supported)")
make_colon_as_title(sections)
bull = bullets_category(
[t for t in random_choices([t for t, _ in sections], k=100)])
if bull >= 0:
chunks = ["\n".join(ck)
for ck in hierarchical_merge(bull, sections, 5)]
else:
sections = [s.split("@") for s, _ in sections]
sections = [(pr[0], "@" + pr[1]) if len(pr) == 2 else (pr[0], '') for pr in sections ]
chunks = naive_merge(
sections, kwargs.get(
"chunk_token_num", 256), kwargs.get(
"delimer", "\n。；！？"))
# is it English
# is_english(random_choices([t for t, _ in sections], k=218))
eng = lang.lower() == "english"
res = tokenize_table(tbls, doc, eng)
res.extend(tokenize_chunks(chunks, doc, eng, pdf_parser))
return res
if __name__ == "__main__":
import sys
def dummy(prog=None, msg=""):
pass
chunk(sys.argv[1], from_page=1, to_page=10, callback=dummy)
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
#
import copy
from tika import parser
import re
from io import BytesIO
from rag.nlp import bullets_category, is_english, tokenize, remove_contents_table, \
hierarchical_merge, make_colon_as_title, naive_merge, random_choices, tokenize_table, add_positions, \
tokenize_chunks, find_codec
from rag.nlp import rag_tokenizer
from deepdoc.parser import PdfParser, DocxParser, PlainParser, HtmlParser
class Pdf(PdfParser):
def __call__(self, filename, binary=None, from_page=0,
to_page=100000, zoomin=3, callback=None):
callback(msg="OCR is running...")
self.__images__(
filename if not binary else binary,
zoomin,
from_page,
to_page,
callback)
callback(msg="OCR finished")
from timeit import default_timer as timer
start = timer()
self._layouts_rec(zoomin)
callback(0.67, "Layout analysis finished")
print("layouts:", timer() - start)
self._table_transformer_job(zoomin)
callback(0.68, "Table analysis finished")
self._text_merge()
tbls = self._extract_table_figure(True, zoomin, True, True)
self._naive_vertical_merge()
self._filter_forpages()
self._merge_with_same_bullet()
callback(0.75, "Text merging finished.")
callback(0.8, "Text extraction finished")
return [(b["text"] + self._line_tag(b, zoomin), b.get("layoutno", ""))
for b in self.boxes], tbls
def chunk(filename, binary=None, from_page=0, to_page=100000,
lang="Chinese", callback=None, **kwargs):
"""
Supported file formats are docx, pdf, txt.
Since a book is long and not all the parts are useful, if it's a PDF,
please setup the page ranges for every book in order eliminate negative effects and save elapsed computing time.
"""
doc = {
"docnm_kwd": filename,
"title_tks": rag_tokenizer.tokenize(re.sub(r"\.[a-zA-Z]+$", "", filename))
}
doc["title_sm_tks"] = rag_tokenizer.fine_grained_tokenize(doc["title_tks"])
pdf_parser = None
sections, tbls = [], []
if re.search(r"\.docx$", filename, re.IGNORECASE):
callback(0.1, "Start to parse.")
doc_parser = DocxParser()
# TODO: table of contents need to be removed
sections, tbls = doc_parser(
binary if binary else filename, from_page=from_page, to_page=to_page)
remove_contents_table(sections, eng=is_english(
random_choices([t for t, _ in sections], k=200)))
tbls = [((None, lns), None) for lns in tbls]
callback(0.8, "Finish parsing.")
elif re.search(r"\.pdf$", filename, re.IGNORECASE):
pdf_parser = Pdf() if kwargs.get(
"parser_config", {}).get(
"layout_recognize", True) else PlainParser()
sections, tbls = pdf_parser(filename if not binary else binary,
from_page=from_page, to_page=to_page, callback=callback)
elif re.search(r"\.txt$", filename, re.IGNORECASE):
callback(0.1, "Start to parse.")
txt = ""
if binary:
encoding = find_codec(binary)
txt = binary.decode(encoding, errors="ignore")
else:
with open(filename, "r") as f:
while True:
l = f.readline()
if not l:
break
txt += l
sections = txt.split("\n")
sections = [(l, "") for l in sections if l]
remove_contents_table(sections, eng=is_english(
random_choices([t for t, _ in sections], k=200)))
callback(0.8, "Finish parsing.")
elif re.search(r"\.(htm|html)$", filename, re.IGNORECASE):
callback(0.1, "Start to parse.")
sections = HtmlParser()(filename, binary)
sections = [(l, "") for l in sections if l]
remove_contents_table(sections, eng=is_english(
random_choices([t for t, _ in sections], k=200)))
callback(0.8, "Finish parsing.")
elif re.search(r"\.doc$", filename, re.IGNORECASE):
callback(0.1, "Start to parse.")
binary = BytesIO(binary)
doc_parsed = parser.from_buffer(binary)
sections = doc_parsed['content'].split('\n')
sections = [(l, "") for l in sections if l]
remove_contents_table(sections, eng=is_english(
random_choices([t for t, _ in sections], k=200)))
callback(0.8, "Finish parsing.")
else:
raise NotImplementedError(
"file type not supported yet(doc, docx, pdf, txt supported)")
make_colon_as_title(sections)
bull = bullets_category(
[t for t in random_choices([t for t, _ in sections], k=100)])
if bull >= 0:
chunks = ["\n".join(ck)
for ck in hierarchical_merge(bull, sections, 5)]
else:
sections = [s.split("@") for s, _ in sections]
sections = [(pr[0], "@" + pr[1]) if len(pr) == 2 else (pr[0], '') for pr in sections ]
chunks = naive_merge(
sections, kwargs.get(
"chunk_token_num", 256), kwargs.get(
"delimer", "\n。；！？"))
# is it English
# is_english(random_choices([t for t, _ in sections], k=218))
eng = lang.lower() == "english"
res = tokenize_table(tbls, doc, eng)
res.extend(tokenize_chunks(chunks, doc, eng, pdf_parser))
return res
if __name__ == "__main__":
import sys
def dummy(prog=None, msg=""):
pass
chunk(sys.argv[1], from_page=1, to_page=10, callback=dummy)

440
rag/app/laws.py
View File

@ -1,220 +1,220 @@
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
#
import copy
from tika import parser
import re
from io import BytesIO
from docx import Document
from api.db import ParserType
from rag.nlp import bullets_category, is_english, tokenize, remove_contents_table, hierarchical_merge, \
make_colon_as_title, add_positions, tokenize_chunks, find_codec, docx_question_level
from rag.nlp import rag_tokenizer
from deepdoc.parser import PdfParser, DocxParser, PlainParser, HtmlParser
from rag.settings import cron_logger
class Docx(DocxParser):
def __init__(self):
pass
def __clean(self, line):
line = re.sub(r"\u3000", " ", line).strip()
return line
def old_call(self, filename, binary=None, from_page=0, to_page=100000):
self.doc = Document(
filename) if not binary else Document(BytesIO(binary))
pn = 0
lines = []
for p in self.doc.paragraphs:
if pn > to_page:
break
if from_page <= pn < to_page and p.text.strip():
lines.append(self.__clean(p.text))
for run in p.runs:
if 'lastRenderedPageBreak' in run._element.xml:
pn += 1
continue
if 'w:br' in run._element.xml and 'type="page"' in run._element.xml:
pn += 1
return [l for l in lines if l]
def __call__(self, filename, binary=None, from_page=0, to_page=100000):
self.doc = Document(
filename) if not binary else Document(BytesIO(binary))
pn = 0
lines = []
bull = bullets_category([p.text for p in self.doc.paragraphs])
for p in self.doc.paragraphs:
if pn > to_page:
break
question_level, p_text = docx_question_level(p, bull)
if not p_text.strip("\n"):continue
lines.append((question_level, p_text))
for run in p.runs:
if 'lastRenderedPageBreak' in run._element.xml:
pn += 1
continue
if 'w:br' in run._element.xml and 'type="page"' in run._element.xml:
pn += 1
visit = [False for _ in range(len(lines))]
sections = []
for s in range(len(lines)):
e = s + 1
while e < len(lines):
if lines[e][0] <= lines[s][0]:
break
e += 1
if e - s == 1 and visit[s]: continue
sec = []
next_level = lines[s][0] + 1
while not sec and next_level < 22:
for i in range(s+1, e):
if lines[i][0] != next_level: continue
sec.append(lines[i][1])
visit[i] = True
next_level += 1
sec.insert(0, lines[s][1])
sections.append("\n".join(sec))
return [l for l in sections if l]
def __str__(self) -> str:
return f'''
question:{self.question},
answer:{self.answer},
level:{self.level},
childs:{self.childs}
'''
class Pdf(PdfParser):
def __init__(self):
self.model_speciess = ParserType.LAWS.value
super().__init__()
def __call__(self, filename, binary=None, from_page=0,
to_page=100000, zoomin=3, callback=None):
callback(msg="OCR is running...")
self.__images__(
filename if not binary else binary,
zoomin,
from_page,
to_page,
callback
)
callback(msg="OCR finished")
from timeit import default_timer as timer
start = timer()
self._layouts_rec(zoomin)
callback(0.67, "Layout analysis finished")
cron_logger.info("layouts:".format(
(timer() - start) / (self.total_page + 0.1)))
self._naive_vertical_merge()
callback(0.8, "Text extraction finished")
return [(b["text"], self._line_tag(b, zoomin))
for b in self.boxes], None
def chunk(filename, binary=None, from_page=0, to_page=100000,
lang="Chinese", callback=None, **kwargs):
"""
Supported file formats are docx, pdf, txt.
"""
doc = {
"docnm_kwd": filename,
"title_tks": rag_tokenizer.tokenize(re.sub(r"\.[a-zA-Z]+$", "", filename))
}
doc["title_sm_tks"] = rag_tokenizer.fine_grained_tokenize(doc["title_tks"])
pdf_parser = None
sections = []
# is it English
eng = lang.lower() == "english" # is_english(sections)
if re.search(r"\.docx$", filename, re.IGNORECASE):
callback(0.1, "Start to parse.")
for txt in Docx()(filename, binary):
sections.append(txt)
callback(0.8, "Finish parsing.")
chunks = sections
return tokenize_chunks(chunks, doc, eng, pdf_parser)
elif re.search(r"\.pdf$", filename, re.IGNORECASE):
pdf_parser = Pdf() if kwargs.get(
"parser_config", {}).get(
"layout_recognize", True) else PlainParser()
for txt, poss in pdf_parser(filename if not binary else binary,
from_page=from_page, to_page=to_page, callback=callback)[0]:
sections.append(txt + poss)
elif re.search(r"\.txt$", filename, re.IGNORECASE):
callback(0.1, "Start to parse.")
txt = ""
if binary:
encoding = find_codec(binary)
txt = binary.decode(encoding, errors="ignore")
else:
with open(filename, "r") as f:
while True:
l = f.readline()
if not l:
break
txt += l
sections = txt.split("\n")
sections = [l for l in sections if l]
callback(0.8, "Finish parsing.")
elif re.search(r"\.(htm|html)$", filename, re.IGNORECASE):
callback(0.1, "Start to parse.")
sections = HtmlParser()(filename, binary)
sections = [l for l in sections if l]
callback(0.8, "Finish parsing.")
elif re.search(r"\.doc$", filename, re.IGNORECASE):
callback(0.1, "Start to parse.")
binary = BytesIO(binary)
doc_parsed = parser.from_buffer(binary)
sections = doc_parsed['content'].split('\n')
sections = [l for l in sections if l]
callback(0.8, "Finish parsing.")
else:
raise NotImplementedError(
"file type not supported yet(doc, docx, pdf, txt supported)")
# Remove 'Contents' part
remove_contents_table(sections, eng)
make_colon_as_title(sections)
bull = bullets_category(sections)
chunks = hierarchical_merge(bull, sections, 5)
if not chunks:
callback(0.99, "No chunk parsed out.")
return tokenize_chunks(["\n".join(ck)
for ck in chunks], doc, eng, pdf_parser)
if __name__ == "__main__":
import sys
def dummy(prog=None, msg=""):
pass
chunk(sys.argv[1], callback=dummy)
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
#
import copy
from tika import parser
import re
from io import BytesIO
from docx import Document
from api.db import ParserType
from rag.nlp import bullets_category, is_english, tokenize, remove_contents_table, hierarchical_merge, \
make_colon_as_title, add_positions, tokenize_chunks, find_codec, docx_question_level
from rag.nlp import rag_tokenizer
from deepdoc.parser import PdfParser, DocxParser, PlainParser, HtmlParser
from rag.settings import cron_logger
class Docx(DocxParser):
def __init__(self):
pass
def __clean(self, line):
line = re.sub(r"\u3000", " ", line).strip()
return line
def old_call(self, filename, binary=None, from_page=0, to_page=100000):
self.doc = Document(
filename) if not binary else Document(BytesIO(binary))
pn = 0
lines = []
for p in self.doc.paragraphs:
if pn > to_page:
break
if from_page <= pn < to_page and p.text.strip():
lines.append(self.__clean(p.text))
for run in p.runs:
if 'lastRenderedPageBreak' in run._element.xml:
pn += 1
continue
if 'w:br' in run._element.xml and 'type="page"' in run._element.xml:
pn += 1
return [l for l in lines if l]
def __call__(self, filename, binary=None, from_page=0, to_page=100000):
self.doc = Document(
filename) if not binary else Document(BytesIO(binary))
pn = 0
lines = []
bull = bullets_category([p.text for p in self.doc.paragraphs])
for p in self.doc.paragraphs:
if pn > to_page:
break
question_level, p_text = docx_question_level(p, bull)
if not p_text.strip("\n"):continue
lines.append((question_level, p_text))
for run in p.runs:
if 'lastRenderedPageBreak' in run._element.xml:
pn += 1
continue
if 'w:br' in run._element.xml and 'type="page"' in run._element.xml:
pn += 1
visit = [False for _ in range(len(lines))]
sections = []
for s in range(len(lines)):
e = s + 1
while e < len(lines):
if lines[e][0] <= lines[s][0]:
break
e += 1
if e - s == 1 and visit[s]: continue
sec = []
next_level = lines[s][0] + 1
while not sec and next_level < 22:
for i in range(s+1, e):
if lines[i][0] != next_level: continue
sec.append(lines[i][1])
visit[i] = True
next_level += 1
sec.insert(0, lines[s][1])
sections.append("\n".join(sec))
return [l for l in sections if l]
def __str__(self) -> str:
return f'''
question:{self.question},
answer:{self.answer},
level:{self.level},
childs:{self.childs}
'''
class Pdf(PdfParser):
def __init__(self):
self.model_speciess = ParserType.LAWS.value
super().__init__()
def __call__(self, filename, binary=None, from_page=0,
to_page=100000, zoomin=3, callback=None):
callback(msg="OCR is running...")
self.__images__(
filename if not binary else binary,
zoomin,
from_page,
to_page,
callback
)
callback(msg="OCR finished")
from timeit import default_timer as timer
start = timer()
self._layouts_rec(zoomin)
callback(0.67, "Layout analysis finished")
cron_logger.info("layouts:".format(
(timer() - start) / (self.total_page + 0.1)))
self._naive_vertical_merge()
callback(0.8, "Text extraction finished")
return [(b["text"], self._line_tag(b, zoomin))
for b in self.boxes], None
def chunk(filename, binary=None, from_page=0, to_page=100000,
lang="Chinese", callback=None, **kwargs):
"""
Supported file formats are docx, pdf, txt.
"""
doc = {
"docnm_kwd": filename,
"title_tks": rag_tokenizer.tokenize(re.sub(r"\.[a-zA-Z]+$", "", filename))
}
doc["title_sm_tks"] = rag_tokenizer.fine_grained_tokenize(doc["title_tks"])
pdf_parser = None
sections = []
# is it English
eng = lang.lower() == "english" # is_english(sections)
if re.search(r"\.docx$", filename, re.IGNORECASE):
callback(0.1, "Start to parse.")
for txt in Docx()(filename, binary):
sections.append(txt)
callback(0.8, "Finish parsing.")
chunks = sections
return tokenize_chunks(chunks, doc, eng, pdf_parser)
elif re.search(r"\.pdf$", filename, re.IGNORECASE):
pdf_parser = Pdf() if kwargs.get(
"parser_config", {}).get(
"layout_recognize", True) else PlainParser()
for txt, poss in pdf_parser(filename if not binary else binary,
from_page=from_page, to_page=to_page, callback=callback)[0]:
sections.append(txt + poss)
elif re.search(r"\.txt$", filename, re.IGNORECASE):
callback(0.1, "Start to parse.")
txt = ""
if binary:
encoding = find_codec(binary)
txt = binary.decode(encoding, errors="ignore")
else:
with open(filename, "r") as f:
while True:
l = f.readline()
if not l:
break
txt += l
sections = txt.split("\n")
sections = [l for l in sections if l]
callback(0.8, "Finish parsing.")
elif re.search(r"\.(htm|html)$", filename, re.IGNORECASE):
callback(0.1, "Start to parse.")
sections = HtmlParser()(filename, binary)
sections = [l for l in sections if l]
callback(0.8, "Finish parsing.")
elif re.search(r"\.doc$", filename, re.IGNORECASE):
callback(0.1, "Start to parse.")
binary = BytesIO(binary)
doc_parsed = parser.from_buffer(binary)
sections = doc_parsed['content'].split('\n')
sections = [l for l in sections if l]
callback(0.8, "Finish parsing.")
else:
raise NotImplementedError(
"file type not supported yet(doc, docx, pdf, txt supported)")
# Remove 'Contents' part
remove_contents_table(sections, eng)
make_colon_as_title(sections)
bull = bullets_category(sections)
chunks = hierarchical_merge(bull, sections, 5)
if not chunks:
callback(0.99, "No chunk parsed out.")
return tokenize_chunks(["\n".join(ck)
for ck in chunks], doc, eng, pdf_parser)
if __name__ == "__main__":
import sys
def dummy(prog=None, msg=""):
pass
chunk(sys.argv[1], callback=dummy)

542
rag/app/manual.py
View File

@ -1,272 +1,272 @@
#
# Copyright 2024 The InfiniFlow Authors. All Rights Reserved.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
#
import copy
import re
from api.db import ParserType
from io import BytesIO
from rag.nlp import rag_tokenizer, tokenize, tokenize_table, add_positions, bullets_category, title_frequency, tokenize_chunks, docx_question_level
from deepdoc.parser import PdfParser, PlainParser
from rag.utils import num_tokens_from_string
from deepdoc.parser import PdfParser, ExcelParser, DocxParser
from docx import Document
from PIL import Image
class Pdf(PdfParser):
def __init__(self):
self.model_speciess = ParserType.MANUAL.value
super().__init__()
def __call__(self, filename, binary=None, from_page=0,
to_page=100000, zoomin=3, callback=None):
from timeit import default_timer as timer
start = timer()
callback(msg="OCR is running...")
self.__images__(
filename if not binary else binary,
zoomin,
from_page,
to_page,
callback
)
callback(msg="OCR finished.")
# for bb in self.boxes:
# for b in bb:
# print(b)
print("OCR:", timer() - start)
self._layouts_rec(zoomin)
callback(0.65, "Layout analysis finished.")
print("layouts:", timer() - start)
self._table_transformer_job(zoomin)
callback(0.67, "Table analysis finished.")
self._text_merge()
tbls = self._extract_table_figure(True, zoomin, True, True)
self._concat_downward()
self._filter_forpages()
callback(0.68, "Text merging finished")
# clean mess
for b in self.boxes:
b["text"] = re.sub(r"([\t 　]|\u3000){2,}", " ", b["text"].strip())
return [(b["text"], b.get("layout_no", ""), self.get_position(b, zoomin))
for i, b in enumerate(self.boxes)], tbls
class Docx(DocxParser):
def __init__(self):
pass
def get_picture(self, document, paragraph):
img = paragraph._element.xpath('.//pic:pic')
if not img:
return None
img = img[0]
embed = img.xpath('.//a:blip/@r:embed')[0]
related_part = document.part.related_parts[embed]
image = related_part.image
image = Image.open(BytesIO(image.blob))
return image
def concat_img(self, img1, img2):
if img1 and not img2:
return img1
if not img1 and img2:
return img2
if not img1 and not img2:
return None
width1, height1 = img1.size
width2, height2 = img2.size
new_width = max(width1, width2)
new_height = height1 + height2
new_image = Image.new('RGB', (new_width, new_height))
new_image.paste(img1, (0, 0))
new_image.paste(img2, (0, height1))
return new_image
def __call__(self, filename, binary=None, from_page=0, to_page=100000, callback=None):
self.doc = Document(
filename) if not binary else Document(BytesIO(binary))
pn = 0
last_answer, last_image = "", None
question_stack, level_stack = [], []
ti_list = []
for p in self.doc.paragraphs:
if pn > to_page:
break
question_level, p_text = 0, ''
if from_page <= pn < to_page and p.text.strip():
question_level, p_text = docx_question_level(p)
if not question_level or question_level > 6: # not a question
last_answer = f'{last_answer}\n{p_text}'
current_image = self.get_picture(self.doc, p)
last_image = self.concat_img(last_image, current_image)
else: # is a question
if last_answer or last_image:
sum_question = '\n'.join(question_stack)
if sum_question:
ti_list.append((f'{sum_question}\n{last_answer}', last_image))
last_answer, last_image = '', None
i = question_level
while question_stack and i <= level_stack[-1]:
question_stack.pop()
level_stack.pop()
question_stack.append(p_text)
level_stack.append(question_level)
for run in p.runs:
if 'lastRenderedPageBreak' in run._element.xml:
pn += 1
continue
if 'w:br' in run._element.xml and 'type="page"' in run._element.xml:
pn += 1
if last_answer:
sum_question = '\n'.join(question_stack)
if sum_question:
ti_list.append((f'{sum_question}\n{last_answer}', last_image))
tbls = []
for tb in self.doc.tables:
html= "<table>"
for r in tb.rows:
html += "<tr>"
i = 0
while i < len(r.cells):
span = 1
c = r.cells[i]
for j in range(i+1, len(r.cells)):
if c.text == r.cells[j].text:
span += 1
i = j
i += 1
html += f"<td>{c.text}</td>" if span == 1 else f"<td colspan='{span}'>{c.text}</td>"
html += "</tr>"
html += "</table>"
tbls.append(((None, html), ""))
return ti_list, tbls
def chunk(filename, binary=None, from_page=0, to_page=100000,
lang="Chinese", callback=None, **kwargs):
"""
Only pdf is supported.
"""
pdf_parser = None
doc = {
"docnm_kwd": filename
}
doc["title_tks"] = rag_tokenizer.tokenize(re.sub(r"\.[a-zA-Z]+$", "", doc["docnm_kwd"]))
doc["title_sm_tks"] = rag_tokenizer.fine_grained_tokenize(doc["title_tks"])
# is it English
eng = lang.lower() == "english" # pdf_parser.is_english
if re.search(r"\.pdf$", filename, re.IGNORECASE):
pdf_parser = Pdf() if kwargs.get(
"parser_config", {}).get(
"layout_recognize", True) else PlainParser()
sections, tbls = pdf_parser(filename if not binary else binary,
from_page=from_page, to_page=to_page, callback=callback)
if sections and len(sections[0]) < 3:
sections = [(t, l, [[0] * 5]) for t, l in sections]
# set pivot using the most frequent type of title,
# then merge between 2 pivot
if len(sections) > 0 and len(pdf_parser.outlines) / len(sections) > 0.1:
max_lvl = max([lvl for _, lvl in pdf_parser.outlines])
most_level = max(0, max_lvl - 1)
levels = []
for txt, _, _ in sections:
for t, lvl in pdf_parser.outlines:
tks = set([t[i] + t[i + 1] for i in range(len(t) - 1)])
tks_ = set([txt[i] + txt[i + 1]
for i in range(min(len(t), len(txt) - 1))])
if len(set(tks & tks_)) / max([len(tks), len(tks_), 1]) > 0.8:
levels.append(lvl)
break
else:
levels.append(max_lvl + 1)
else:
bull = bullets_category([txt for txt, _, _ in sections])
most_level, levels = title_frequency(
bull, [(txt, l) for txt, l, poss in sections])
assert len(sections) == len(levels)
sec_ids = []
sid = 0
for i, lvl in enumerate(levels):
if lvl <= most_level and i > 0 and lvl != levels[i - 1]:
sid += 1
sec_ids.append(sid)
# print(lvl, self.boxes[i]["text"], most_level, sid)
sections = [(txt, sec_ids[i], poss)
for i, (txt, _, poss) in enumerate(sections)]
for (img, rows), poss in tbls:
if not rows: continue
sections.append((rows if isinstance(rows, str) else rows[0], -1,
[(p[0] + 1 - from_page, p[1], p[2], p[3], p[4]) for p in poss]))
def tag(pn, left, right, top, bottom):
if pn + left + right + top + bottom == 0:
return ""
return "@@{}\t{:.1f}\t{:.1f}\t{:.1f}\t{:.1f}##" \
.format(pn, left, right, top, bottom)
chunks = []
last_sid = -2
tk_cnt = 0
for txt, sec_id, poss in sorted(sections, key=lambda x: (
x[-1][0][0], x[-1][0][3], x[-1][0][1])):
poss = "\t".join([tag(*pos) for pos in poss])
if tk_cnt < 32 or (tk_cnt < 1024 and (sec_id == last_sid or sec_id == -1)):
if chunks:
chunks[-1] += "\n" + txt + poss
tk_cnt += num_tokens_from_string(txt)
continue
chunks.append(txt + poss)
tk_cnt = num_tokens_from_string(txt)
if sec_id > -1:
last_sid = sec_id
res = tokenize_table(tbls, doc, eng)
res.extend(tokenize_chunks(chunks, doc, eng, pdf_parser))
return res
if re.search(r"\.docx$", filename, re.IGNORECASE):
docx_parser = Docx()
ti_list, tbls = docx_parser(filename, binary,
from_page=0, to_page=10000, callback=callback)
res = tokenize_table(tbls, doc, eng)
for text, image in ti_list:
d = copy.deepcopy(doc)
d['image'] = image
tokenize(d, text, eng)
res.append(d)
return res
else:
raise NotImplementedError("file type not supported yet(pdf and docx supported)")
if __name__ == "__main__":
import sys
def dummy(prog=None, msg=""):
pass
#
# Copyright 2024 The InfiniFlow Authors. All Rights Reserved.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
#
import copy
import re
from api.db import ParserType
from io import BytesIO
from rag.nlp import rag_tokenizer, tokenize, tokenize_table, add_positions, bullets_category, title_frequency, tokenize_chunks, docx_question_level
from deepdoc.parser import PdfParser, PlainParser
from rag.utils import num_tokens_from_string
from deepdoc.parser import PdfParser, ExcelParser, DocxParser
from docx import Document
from PIL import Image
class Pdf(PdfParser):
def __init__(self):
self.model_speciess = ParserType.MANUAL.value
super().__init__()
def __call__(self, filename, binary=None, from_page=0,
to_page=100000, zoomin=3, callback=None):
from timeit import default_timer as timer
start = timer()
callback(msg="OCR is running...")
self.__images__(
filename if not binary else binary,
zoomin,
from_page,
to_page,
callback
)
callback(msg="OCR finished.")
# for bb in self.boxes:
# for b in bb:
# print(b)
print("OCR:", timer() - start)
self._layouts_rec(zoomin)
callback(0.65, "Layout analysis finished.")
print("layouts:", timer() - start)
self._table_transformer_job(zoomin)
callback(0.67, "Table analysis finished.")
self._text_merge()
tbls = self._extract_table_figure(True, zoomin, True, True)
self._concat_downward()
self._filter_forpages()
callback(0.68, "Text merging finished")
# clean mess
for b in self.boxes:
b["text"] = re.sub(r"([\t 　]|\u3000){2,}", " ", b["text"].strip())
return [(b["text"], b.get("layout_no", ""), self.get_position(b, zoomin))
for i, b in enumerate(self.boxes)], tbls
class Docx(DocxParser):
def __init__(self):
pass
def get_picture(self, document, paragraph):
img = paragraph._element.xpath('.//pic:pic')
if not img:
return None
img = img[0]
embed = img.xpath('.//a:blip/@r:embed')[0]
related_part = document.part.related_parts[embed]
image = related_part.image
image = Image.open(BytesIO(image.blob))
return image
def concat_img(self, img1, img2):
if img1 and not img2:
return img1
if not img1 and img2:
return img2
if not img1 and not img2:
return None
width1, height1 = img1.size
width2, height2 = img2.size
new_width = max(width1, width2)
new_height = height1 + height2
new_image = Image.new('RGB', (new_width, new_height))
new_image.paste(img1, (0, 0))
new_image.paste(img2, (0, height1))
return new_image
def __call__(self, filename, binary=None, from_page=0, to_page=100000, callback=None):
self.doc = Document(
filename) if not binary else Document(BytesIO(binary))
pn = 0
last_answer, last_image = "", None
question_stack, level_stack = [], []
ti_list = []
for p in self.doc.paragraphs:
if pn > to_page:
break
question_level, p_text = 0, ''
if from_page <= pn < to_page and p.text.strip():
question_level, p_text = docx_question_level(p)
if not question_level or question_level > 6: # not a question
last_answer = f'{last_answer}\n{p_text}'
current_image = self.get_picture(self.doc, p)
last_image = self.concat_img(last_image, current_image)
else: # is a question
if last_answer or last_image:
sum_question = '\n'.join(question_stack)
if sum_question:
ti_list.append((f'{sum_question}\n{last_answer}', last_image))
last_answer, last_image = '', None
i = question_level
while question_stack and i <= level_stack[-1]:
question_stack.pop()
level_stack.pop()
question_stack.append(p_text)
level_stack.append(question_level)
for run in p.runs:
if 'lastRenderedPageBreak' in run._element.xml:
pn += 1
continue
if 'w:br' in run._element.xml and 'type="page"' in run._element.xml:
pn += 1
if last_answer:
sum_question = '\n'.join(question_stack)
if sum_question:
ti_list.append((f'{sum_question}\n{last_answer}', last_image))
tbls = []
for tb in self.doc.tables:
html= "<table>"
for r in tb.rows:
html += "<tr>"
i = 0
while i < len(r.cells):
span = 1
c = r.cells[i]
for j in range(i+1, len(r.cells)):
if c.text == r.cells[j].text:
span += 1
i = j
i += 1
html += f"<td>{c.text}</td>" if span == 1 else f"<td colspan='{span}'>{c.text}</td>"
html += "</tr>"
html += "</table>"
tbls.append(((None, html), ""))
return ti_list, tbls
def chunk(filename, binary=None, from_page=0, to_page=100000,
lang="Chinese", callback=None, **kwargs):
"""
Only pdf is supported.
"""
pdf_parser = None
doc = {
"docnm_kwd": filename
}
doc["title_tks"] = rag_tokenizer.tokenize(re.sub(r"\.[a-zA-Z]+$", "", doc["docnm_kwd"]))
doc["title_sm_tks"] = rag_tokenizer.fine_grained_tokenize(doc["title_tks"])
# is it English
eng = lang.lower() == "english" # pdf_parser.is_english
if re.search(r"\.pdf$", filename, re.IGNORECASE):
pdf_parser = Pdf() if kwargs.get(
"parser_config", {}).get(
"layout_recognize", True) else PlainParser()
sections, tbls = pdf_parser(filename if not binary else binary,
from_page=from_page, to_page=to_page, callback=callback)
if sections and len(sections[0]) < 3:
sections = [(t, l, [[0] * 5]) for t, l in sections]
# set pivot using the most frequent type of title,
# then merge between 2 pivot
if len(sections) > 0 and len(pdf_parser.outlines) / len(sections) > 0.1:
max_lvl = max([lvl for _, lvl in pdf_parser.outlines])
most_level = max(0, max_lvl - 1)
levels = []
for txt, _, _ in sections:
for t, lvl in pdf_parser.outlines:
tks = set([t[i] + t[i + 1] for i in range(len(t) - 1)])
tks_ = set([txt[i] + txt[i + 1]
for i in range(min(len(t), len(txt) - 1))])
if len(set(tks & tks_)) / max([len(tks), len(tks_), 1]) > 0.8:
levels.append(lvl)
break
else:
levels.append(max_lvl + 1)
else:
bull = bullets_category([txt for txt, _, _ in sections])
most_level, levels = title_frequency(
bull, [(txt, l) for txt, l, poss in sections])
assert len(sections) == len(levels)
sec_ids = []
sid = 0
for i, lvl in enumerate(levels):
if lvl <= most_level and i > 0 and lvl != levels[i - 1]:
sid += 1
sec_ids.append(sid)
# print(lvl, self.boxes[i]["text"], most_level, sid)
sections = [(txt, sec_ids[i], poss)
for i, (txt, _, poss) in enumerate(sections)]
for (img, rows), poss in tbls:
if not rows: continue
sections.append((rows if isinstance(rows, str) else rows[0], -1,
[(p[0] + 1 - from_page, p[1], p[2], p[3], p[4]) for p in poss]))
def tag(pn, left, right, top, bottom):
if pn + left + right + top + bottom == 0:
return ""
return "@@{}\t{:.1f}\t{:.1f}\t{:.1f}\t{:.1f}##" \
.format(pn, left, right, top, bottom)
chunks = []
last_sid = -2
tk_cnt = 0
for txt, sec_id, poss in sorted(sections, key=lambda x: (
x[-1][0][0], x[-1][0][3], x[-1][0][1])):
poss = "\t".join([tag(*pos) for pos in poss])
if tk_cnt < 32 or (tk_cnt < 1024 and (sec_id == last_sid or sec_id == -1)):
if chunks:
chunks[-1] += "\n" + txt + poss
tk_cnt += num_tokens_from_string(txt)
continue
chunks.append(txt + poss)
tk_cnt = num_tokens_from_string(txt)
if sec_id > -1:
last_sid = sec_id
res = tokenize_table(tbls, doc, eng)
res.extend(tokenize_chunks(chunks, doc, eng, pdf_parser))
return res
if re.search(r"\.docx$", filename, re.IGNORECASE):
docx_parser = Docx()
ti_list, tbls = docx_parser(filename, binary,
from_page=0, to_page=10000, callback=callback)
res = tokenize_table(tbls, doc, eng)
for text, image in ti_list:
d = copy.deepcopy(doc)
d['image'] = image
tokenize(d, text, eng)
res.append(d)
return res
else:
raise NotImplementedError("file type not supported yet(pdf and docx supported)")
if __name__ == "__main__":
import sys
def dummy(prog=None, msg=""):
pass
chunk(sys.argv[1], callback=dummy)

564
rag/app/naive.py
View File

@ -1,282 +1,282 @@
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
#
from tika import parser
from io import BytesIO
from docx import Document
from timeit import default_timer as timer
import re
from deepdoc.parser.pdf_parser import PlainParser
from rag.nlp import rag_tokenizer, naive_merge, tokenize_table, tokenize_chunks, find_codec, concat_img, naive_merge_docx, tokenize_chunks_docx
from deepdoc.parser import PdfParser, ExcelParser, DocxParser, HtmlParser, JsonParser, MarkdownParser, TxtParser
from rag.settings import cron_logger
from rag.utils import num_tokens_from_string
from PIL import Image
from functools import reduce
from markdown import markdown
from docx.image.exceptions import UnrecognizedImageError
class Docx(DocxParser):
def __init__(self):
pass
def get_picture(self, document, paragraph):
img = paragraph._element.xpath('.//pic:pic')
if not img:
return None
img = img[0]
embed = img.xpath('.//a:blip/@r:embed')[0]
related_part = document.part.related_parts[embed]
try:
image_blob = related_part.image.blob
except UnrecognizedImageError:
print("Unrecognized image format. Skipping image.")
return None
try:
image = Image.open(BytesIO(image_blob)).convert('RGB')
return image
except Exception as e:
return None
def __clean(self, line):
line = re.sub(r"\u3000", " ", line).strip()
return line
def __call__(self, filename, binary=None, from_page=0, to_page=100000):
self.doc = Document(
filename) if not binary else Document(BytesIO(binary))
pn = 0
lines = []
last_image = None
for p in self.doc.paragraphs:
if pn > to_page:
break
if from_page <= pn < to_page:
if p.text.strip():
if p.style and p.style.name == 'Caption':
former_image = None
if lines and lines[-1][1] and lines[-1][2] != 'Caption':
former_image = lines[-1][1].pop()
elif last_image:
former_image = last_image
last_image = None
lines.append((self.__clean(p.text), [former_image], p.style.name))
else:
current_image = self.get_picture(self.doc, p)
image_list = [current_image]
if last_image:
image_list.insert(0, last_image)
last_image = None
lines.append((self.__clean(p.text), image_list, p.style.name))
else:
if current_image := self.get_picture(self.doc, p):
if lines:
lines[-1][1].append(current_image)
else:
last_image = current_image
for run in p.runs:
if 'lastRenderedPageBreak' in run._element.xml:
pn += 1
continue
if 'w:br' in run._element.xml and 'type="page"' in run._element.xml:
pn += 1
new_line = [(line[0], reduce(concat_img, line[1]) if line[1] else None) for line in lines]
tbls = []
for tb in self.doc.tables:
html= "<table>"
for r in tb.rows:
html += "<tr>"
i = 0
while i < len(r.cells):
span = 1
c = r.cells[i]
for j in range(i+1, len(r.cells)):
if c.text == r.cells[j].text:
span += 1
i = j
i += 1
html += f"<td>{c.text}</td>" if span == 1 else f"<td colspan='{span}'>{c.text}</td>"
html += "</tr>"
html += "</table>"
tbls.append(((None, html), ""))
return new_line, tbls
class Pdf(PdfParser):
def __call__(self, filename, binary=None, from_page=0,
to_page=100000, zoomin=3, callback=None):
start = timer()
callback(msg="OCR is running...")
self.__images__(
filename if not binary else binary,
zoomin,
from_page,
to_page,
callback
)
callback(msg="OCR finished")
cron_logger.info("OCR({}~{}): {}".format(from_page, to_page, timer() - start))
start = timer()
self._layouts_rec(zoomin)
callback(0.63, "Layout analysis finished.")
self._table_transformer_job(zoomin)
callback(0.65, "Table analysis finished.")
self._text_merge()
callback(0.67, "Text merging finished")
tbls = self._extract_table_figure(True, zoomin, True, True)
#self._naive_vertical_merge()
self._concat_downward()
#self._filter_forpages()
cron_logger.info("layouts: {}".format(timer() - start))
return [(b["text"], self._line_tag(b, zoomin))
for b in self.boxes], tbls
class Markdown(MarkdownParser):
def __call__(self, filename, binary=None):
txt = ""
tbls = []
if binary:
encoding = find_codec(binary)
txt = binary.decode(encoding, errors="ignore")
else:
with open(filename, "r") as f:
txt = f.read()
remainder, tables = self.extract_tables_and_remainder(f'{txt}\n')
sections = []
tbls = []
for sec in remainder.split("\n"):
if num_tokens_from_string(sec) > 10 * self.chunk_token_num:
sections.append((sec[:int(len(sec)/2)], ""))
sections.append((sec[int(len(sec)/2):], ""))
else:
sections.append((sec, ""))
print(tables)
for table in tables:
tbls.append(((None, markdown(table, extensions=['markdown.extensions.tables'])), ""))
return sections, tbls
def chunk(filename, binary=None, from_page=0, to_page=100000,
lang="Chinese", callback=None, **kwargs):
"""
Supported file formats are docx, pdf, excel, txt.
This method apply the naive ways to chunk files.
Successive text will be sliced into pieces using 'delimiter'.
Next, these successive pieces are merge into chunks whose token number is no more than 'Max token number'.
"""
eng = lang.lower() == "english" # is_english(cks)
parser_config = kwargs.get(
"parser_config", {
"chunk_token_num": 128, "delimiter": "\n!?。；！？", "layout_recognize": True})
doc = {
"docnm_kwd": filename,
"title_tks": rag_tokenizer.tokenize(re.sub(r"\.[a-zA-Z]+$", "", filename))
}
doc["title_sm_tks"] = rag_tokenizer.fine_grained_tokenize(doc["title_tks"])
res = []
pdf_parser = None
sections = []
if re.search(r"\.docx$", filename, re.IGNORECASE):
callback(0.1, "Start to parse.")
sections, tbls = Docx()(filename, binary)
res = tokenize_table(tbls, doc, eng) # just for table
callback(0.8, "Finish parsing.")
st = timer()
chunks, images = naive_merge_docx(
sections, int(parser_config.get(
"chunk_token_num", 128)), parser_config.get(
"delimiter", "\n!?。；！？"))
if kwargs.get("section_only", False):
return chunks
res.extend(tokenize_chunks_docx(chunks, doc, eng, images))
cron_logger.info("naive_merge({}): {}".format(filename, timer() - st))
return res
elif re.search(r"\.pdf$", filename, re.IGNORECASE):
pdf_parser = Pdf(
) if parser_config.get("layout_recognize", True) else PlainParser()
sections, tbls = pdf_parser(filename if not binary else binary,
from_page=from_page, to_page=to_page, callback=callback)
res = tokenize_table(tbls, doc, eng)
elif re.search(r"\.xlsx?$", filename, re.IGNORECASE):
callback(0.1, "Start to parse.")
excel_parser = ExcelParser()
sections = [(l, "") for l in excel_parser.html(binary) if l]
elif re.search(r"\.(txt|py|js|java|c|cpp|h|php|go|ts|sh|cs|kt|sql)$", filename, re.IGNORECASE):
callback(0.1, "Start to parse.")
sections = TxtParser()(filename,binary,
parser_config.get("chunk_token_num", 128),
parser_config.get("delimiter", "\n!?;。；！？"))
callback(0.8, "Finish parsing.")
elif re.search(r"\.(md|markdown)$", filename, re.IGNORECASE):
callback(0.1, "Start to parse.")
sections, tbls = Markdown(int(parser_config.get("chunk_token_num", 128)))(filename, binary)
res = tokenize_table(tbls, doc, eng)
callback(0.8, "Finish parsing.")
elif re.search(r"\.(htm|html)$", filename, re.IGNORECASE):
callback(0.1, "Start to parse.")
sections = HtmlParser()(filename, binary)
sections = [(l, "") for l in sections if l]
callback(0.8, "Finish parsing.")
elif re.search(r"\.json$", filename, re.IGNORECASE):
callback(0.1, "Start to parse.")
sections = JsonParser(int(parser_config.get("chunk_token_num", 128)))(binary)
sections = [(l, "") for l in sections if l]
callback(0.8, "Finish parsing.")
elif re.search(r"\.doc$", filename, re.IGNORECASE):
callback(0.1, "Start to parse.")
binary = BytesIO(binary)
doc_parsed = parser.from_buffer(binary)
sections = doc_parsed['content'].split('\n')
sections = [(l, "") for l in sections if l]
callback(0.8, "Finish parsing.")
else:
raise NotImplementedError(
"file type not supported yet(pdf, xlsx, doc, docx, txt supported)")
st = timer()
chunks = naive_merge(
sections, int(parser_config.get(
"chunk_token_num", 128)), parser_config.get(
"delimiter", "\n!?。；！？"))
if kwargs.get("section_only", False):
return chunks
res.extend(tokenize_chunks(chunks, doc, eng, pdf_parser))
cron_logger.info("naive_merge({}): {}".format(filename, timer() - st))
return res
if __name__ == "__main__":
import sys
def dummy(prog=None, msg=""):
pass
chunk(sys.argv[1], from_page=0, to_page=10, callback=dummy)
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
#
from tika import parser
from io import BytesIO
from docx import Document
from timeit import default_timer as timer
import re
from deepdoc.parser.pdf_parser import PlainParser
from rag.nlp import rag_tokenizer, naive_merge, tokenize_table, tokenize_chunks, find_codec, concat_img, naive_merge_docx, tokenize_chunks_docx
from deepdoc.parser import PdfParser, ExcelParser, DocxParser, HtmlParser, JsonParser, MarkdownParser, TxtParser
from rag.settings import cron_logger
from rag.utils import num_tokens_from_string
from PIL import Image
from functools import reduce
from markdown import markdown
from docx.image.exceptions import UnrecognizedImageError
class Docx(DocxParser):
def __init__(self):
pass
def get_picture(self, document, paragraph):
img = paragraph._element.xpath('.//pic:pic')
if not img:
return None
img = img[0]
embed = img.xpath('.//a:blip/@r:embed')[0]
related_part = document.part.related_parts[embed]
try:
image_blob = related_part.image.blob
except UnrecognizedImageError:
print("Unrecognized image format. Skipping image.")
return None
try:
image = Image.open(BytesIO(image_blob)).convert('RGB')
return image
except Exception as e:
return None
def __clean(self, line):
line = re.sub(r"\u3000", " ", line).strip()
return line
def __call__(self, filename, binary=None, from_page=0, to_page=100000):
self.doc = Document(
filename) if not binary else Document(BytesIO(binary))
pn = 0
lines = []
last_image = None
for p in self.doc.paragraphs:
if pn > to_page:
break
if from_page <= pn < to_page:
if p.text.strip():
if p.style and p.style.name == 'Caption':
former_image = None
if lines and lines[-1][1] and lines[-1][2] != 'Caption':
former_image = lines[-1][1].pop()
elif last_image:
former_image = last_image
last_image = None
lines.append((self.__clean(p.text), [former_image], p.style.name))
else:
current_image = self.get_picture(self.doc, p)
image_list = [current_image]
if last_image:
image_list.insert(0, last_image)
last_image = None
lines.append((self.__clean(p.text), image_list, p.style.name))
else:
if current_image := self.get_picture(self.doc, p):
if lines:
lines[-1][1].append(current_image)
else:
last_image = current_image
for run in p.runs:
if 'lastRenderedPageBreak' in run._element.xml:
pn += 1
continue
if 'w:br' in run._element.xml and 'type="page"' in run._element.xml:
pn += 1
new_line = [(line[0], reduce(concat_img, line[1]) if line[1] else None) for line in lines]
tbls = []
for tb in self.doc.tables:
html= "<table>"
for r in tb.rows:
html += "<tr>"
i = 0
while i < len(r.cells):
span = 1
c = r.cells[i]
for j in range(i+1, len(r.cells)):
if c.text == r.cells[j].text:
span += 1
i = j
i += 1
html += f"<td>{c.text}</td>" if span == 1 else f"<td colspan='{span}'>{c.text}</td>"
html += "</tr>"
html += "</table>"
tbls.append(((None, html), ""))
return new_line, tbls
class Pdf(PdfParser):
def __call__(self, filename, binary=None, from_page=0,
to_page=100000, zoomin=3, callback=None):
start = timer()
callback(msg="OCR is running...")
self.__images__(
filename if not binary else binary,
zoomin,
from_page,
to_page,
callback
)
callback(msg="OCR finished")
cron_logger.info("OCR({}~{}): {}".format(from_page, to_page, timer() - start))
start = timer()
self._layouts_rec(zoomin)
callback(0.63, "Layout analysis finished.")
self._table_transformer_job(zoomin)
callback(0.65, "Table analysis finished.")
self._text_merge()
callback(0.67, "Text merging finished")
tbls = self._extract_table_figure(True, zoomin, True, True)
#self._naive_vertical_merge()
self._concat_downward()
#self._filter_forpages()
cron_logger.info("layouts: {}".format(timer() - start))
return [(b["text"], self._line_tag(b, zoomin))
for b in self.boxes], tbls
class Markdown(MarkdownParser):
def __call__(self, filename, binary=None):
txt = ""
tbls = []
if binary:
encoding = find_codec(binary)
txt = binary.decode(encoding, errors="ignore")
else:
with open(filename, "r") as f:
txt = f.read()
remainder, tables = self.extract_tables_and_remainder(f'{txt}\n')
sections = []
tbls = []
for sec in remainder.split("\n"):
if num_tokens_from_string(sec) > 10 * self.chunk_token_num:
sections.append((sec[:int(len(sec)/2)], ""))
sections.append((sec[int(len(sec)/2):], ""))
else:
sections.append((sec, ""))
print(tables)
for table in tables:
tbls.append(((None, markdown(table, extensions=['markdown.extensions.tables'])), ""))
return sections, tbls
def chunk(filename, binary=None, from_page=0, to_page=100000,
lang="Chinese", callback=None, **kwargs):
"""
Supported file formats are docx, pdf, excel, txt.
This method apply the naive ways to chunk files.
Successive text will be sliced into pieces using 'delimiter'.
Next, these successive pieces are merge into chunks whose token number is no more than 'Max token number'.
"""
eng = lang.lower() == "english" # is_english(cks)
parser_config = kwargs.get(
"parser_config", {
"chunk_token_num": 128, "delimiter": "\n!?。；！？", "layout_recognize": True})
doc = {
"docnm_kwd": filename,
"title_tks": rag_tokenizer.tokenize(re.sub(r"\.[a-zA-Z]+$", "", filename))
}
doc["title_sm_tks"] = rag_tokenizer.fine_grained_tokenize(doc["title_tks"])
res = []
pdf_parser = None
sections = []
if re.search(r"\.docx$", filename, re.IGNORECASE):
callback(0.1, "Start to parse.")
sections, tbls = Docx()(filename, binary)
res = tokenize_table(tbls, doc, eng) # just for table
callback(0.8, "Finish parsing.")
st = timer()
chunks, images = naive_merge_docx(
sections, int(parser_config.get(
"chunk_token_num", 128)), parser_config.get(
"delimiter", "\n!?。；！？"))
if kwargs.get("section_only", False):
return chunks
res.extend(tokenize_chunks_docx(chunks, doc, eng, images))
cron_logger.info("naive_merge({}): {}".format(filename, timer() - st))
return res
elif re.search(r"\.pdf$", filename, re.IGNORECASE):
pdf_parser = Pdf(
) if parser_config.get("layout_recognize", True) else PlainParser()
sections, tbls = pdf_parser(filename if not binary else binary,
from_page=from_page, to_page=to_page, callback=callback)
res = tokenize_table(tbls, doc, eng)
elif re.search(r"\.xlsx?$", filename, re.IGNORECASE):
callback(0.1, "Start to parse.")
excel_parser = ExcelParser()
sections = [(l, "") for l in excel_parser.html(binary) if l]
elif re.search(r"\.(txt|py|js|java|c|cpp|h|php|go|ts|sh|cs|kt|sql)$", filename, re.IGNORECASE):
callback(0.1, "Start to parse.")
sections = TxtParser()(filename,binary,
parser_config.get("chunk_token_num", 128),
parser_config.get("delimiter", "\n!?;。；！？"))
callback(0.8, "Finish parsing.")
elif re.search(r"\.(md|markdown)$", filename, re.IGNORECASE):
callback(0.1, "Start to parse.")
sections, tbls = Markdown(int(parser_config.get("chunk_token_num", 128)))(filename, binary)
res = tokenize_table(tbls, doc, eng)
callback(0.8, "Finish parsing.")
elif re.search(r"\.(htm|html)$", filename, re.IGNORECASE):
callback(0.1, "Start to parse.")
sections = HtmlParser()(filename, binary)
sections = [(l, "") for l in sections if l]
callback(0.8, "Finish parsing.")
elif re.search(r"\.json$", filename, re.IGNORECASE):
callback(0.1, "Start to parse.")
sections = JsonParser(int(parser_config.get("chunk_token_num", 128)))(binary)
sections = [(l, "") for l in sections if l]
callback(0.8, "Finish parsing.")
elif re.search(r"\.doc$", filename, re.IGNORECASE):
callback(0.1, "Start to parse.")
binary = BytesIO(binary)
doc_parsed = parser.from_buffer(binary)
sections = doc_parsed['content'].split('\n')
sections = [(l, "") for l in sections if l]
callback(0.8, "Finish parsing.")
else:
raise NotImplementedError(
"file type not supported yet(pdf, xlsx, doc, docx, txt supported)")
st = timer()
chunks = naive_merge(
sections, int(parser_config.get(
"chunk_token_num", 128)), parser_config.get(
"delimiter", "\n!?。；！？"))
if kwargs.get("section_only", False):
return chunks
res.extend(tokenize_chunks(chunks, doc, eng, pdf_parser))
cron_logger.info("naive_merge({}): {}".format(filename, timer() - st))
return res
if __name__ == "__main__":
import sys
def dummy(prog=None, msg=""):
pass
chunk(sys.argv[1], from_page=0, to_page=10, callback=dummy)

266
rag/app/one.py
View File

@ -1,133 +1,133 @@
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
#
from tika import parser
from io import BytesIO
import re
from rag.app import laws
from rag.nlp import rag_tokenizer, tokenize, find_codec
from deepdoc.parser import PdfParser, ExcelParser, PlainParser, HtmlParser
class Pdf(PdfParser):
def __call__(self, filename, binary=None, from_page=0,
to_page=100000, zoomin=3, callback=None):
callback(msg="OCR is running...")
self.__images__(
filename if not binary else binary,
zoomin,
from_page,
to_page,
callback
)
callback(msg="OCR finished")
from timeit import default_timer as timer
start = timer()
self._layouts_rec(zoomin, drop=False)
callback(0.63, "Layout analysis finished.")
print("layouts:", timer() - start)
self._table_transformer_job(zoomin)
callback(0.65, "Table analysis finished.")
self._text_merge()
callback(0.67, "Text merging finished")
tbls = self._extract_table_figure(True, zoomin, True, True)
self._concat_downward()
sections = [(b["text"], self.get_position(b, zoomin))
for i, b in enumerate(self.boxes)]
for (img, rows), poss in tbls:
if not rows:continue
sections.append((rows if isinstance(rows, str) else rows[0],
[(p[0] + 1 - from_page, p[1], p[2], p[3], p[4]) for p in poss]))
return [(txt, "") for txt, _ in sorted(sections, key=lambda x: (
x[-1][0][0], x[-1][0][3], x[-1][0][1]))], None
def chunk(filename, binary=None, from_page=0, to_page=100000,
lang="Chinese", callback=None, **kwargs):
"""
Supported file formats are docx, pdf, excel, txt.
One file forms a chunk which maintains original text order.
"""
eng = lang.lower() == "english" # is_english(cks)
if re.search(r"\.docx$", filename, re.IGNORECASE):
callback(0.1, "Start to parse.")
sections = [txt for txt in laws.Docx()(filename, binary) if txt]
callback(0.8, "Finish parsing.")
elif re.search(r"\.pdf$", filename, re.IGNORECASE):
pdf_parser = Pdf() if kwargs.get(
"parser_config", {}).get(
"layout_recognize", True) else PlainParser()
sections, _ = pdf_parser(
filename if not binary else binary, to_page=to_page, callback=callback)
sections = [s for s, _ in sections if s]
elif re.search(r"\.xlsx?$", filename, re.IGNORECASE):
callback(0.1, "Start to parse.")
excel_parser = ExcelParser()
sections = excel_parser.html(binary, 1000000000)
elif re.search(r"\.txt$", filename, re.IGNORECASE):
callback(0.1, "Start to parse.")
txt = ""
if binary:
encoding = find_codec(binary)
txt = binary.decode(encoding, errors="ignore")
else:
with open(filename, "r") as f:
while True:
l = f.readline()
if not l:
break
txt += l
sections = txt.split("\n")
sections = [s for s in sections if s]
callback(0.8, "Finish parsing.")
elif re.search(r"\.(htm|html)$", filename, re.IGNORECASE):
callback(0.1, "Start to parse.")
sections = HtmlParser()(filename, binary)
sections = [s for s in sections if s]
callback(0.8, "Finish parsing.")
elif re.search(r"\.doc$", filename, re.IGNORECASE):
callback(0.1, "Start to parse.")
binary = BytesIO(binary)
doc_parsed = parser.from_buffer(binary)
sections = doc_parsed['content'].split('\n')
sections = [l for l in sections if l]
callback(0.8, "Finish parsing.")
else:
raise NotImplementedError(
"file type not supported yet(doc, docx, pdf, txt supported)")
doc = {
"docnm_kwd": filename,
"title_tks": rag_tokenizer.tokenize(re.sub(r"\.[a-zA-Z]+$", "", filename))
}
doc["title_sm_tks"] = rag_tokenizer.fine_grained_tokenize(doc["title_tks"])
tokenize(doc, "\n".join(sections), eng)
return [doc]
if __name__ == "__main__":
import sys
def dummy(prog=None, msg=""):
pass
chunk(sys.argv[1], from_page=0, to_page=10, callback=dummy)
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
#
from tika import parser
from io import BytesIO
import re
from rag.app import laws
from rag.nlp import rag_tokenizer, tokenize, find_codec
from deepdoc.parser import PdfParser, ExcelParser, PlainParser, HtmlParser
class Pdf(PdfParser):
def __call__(self, filename, binary=None, from_page=0,
to_page=100000, zoomin=3, callback=None):
callback(msg="OCR is running...")
self.__images__(
filename if not binary else binary,
zoomin,
from_page,
to_page,
callback
)
callback(msg="OCR finished")
from timeit import default_timer as timer
start = timer()
self._layouts_rec(zoomin, drop=False)
callback(0.63, "Layout analysis finished.")
print("layouts:", timer() - start)
self._table_transformer_job(zoomin)
callback(0.65, "Table analysis finished.")
self._text_merge()
callback(0.67, "Text merging finished")
tbls = self._extract_table_figure(True, zoomin, True, True)
self._concat_downward()
sections = [(b["text"], self.get_position(b, zoomin))
for i, b in enumerate(self.boxes)]
for (img, rows), poss in tbls:
if not rows:continue
sections.append((rows if isinstance(rows, str) else rows[0],
[(p[0] + 1 - from_page, p[1], p[2], p[3], p[4]) for p in poss]))
return [(txt, "") for txt, _ in sorted(sections, key=lambda x: (
x[-1][0][0], x[-1][0][3], x[-1][0][1]))], None
def chunk(filename, binary=None, from_page=0, to_page=100000,
lang="Chinese", callback=None, **kwargs):
"""
Supported file formats are docx, pdf, excel, txt.
One file forms a chunk which maintains original text order.
"""
eng = lang.lower() == "english" # is_english(cks)
if re.search(r"\.docx$", filename, re.IGNORECASE):
callback(0.1, "Start to parse.")
sections = [txt for txt in laws.Docx()(filename, binary) if txt]
callback(0.8, "Finish parsing.")
elif re.search(r"\.pdf$", filename, re.IGNORECASE):
pdf_parser = Pdf() if kwargs.get(
"parser_config", {}).get(
"layout_recognize", True) else PlainParser()
sections, _ = pdf_parser(
filename if not binary else binary, to_page=to_page, callback=callback)
sections = [s for s, _ in sections if s]
elif re.search(r"\.xlsx?$", filename, re.IGNORECASE):
callback(0.1, "Start to parse.")
excel_parser = ExcelParser()
sections = excel_parser.html(binary, 1000000000)
elif re.search(r"\.txt$", filename, re.IGNORECASE):
callback(0.1, "Start to parse.")
txt = ""
if binary:
encoding = find_codec(binary)
txt = binary.decode(encoding, errors="ignore")
else:
with open(filename, "r") as f:
while True:
l = f.readline()
if not l:
break
txt += l
sections = txt.split("\n")
sections = [s for s in sections if s]
callback(0.8, "Finish parsing.")
elif re.search(r"\.(htm|html)$", filename, re.IGNORECASE):
callback(0.1, "Start to parse.")
sections = HtmlParser()(filename, binary)
sections = [s for s in sections if s]
callback(0.8, "Finish parsing.")
elif re.search(r"\.doc$", filename, re.IGNORECASE):
callback(0.1, "Start to parse.")
binary = BytesIO(binary)
doc_parsed = parser.from_buffer(binary)
sections = doc_parsed['content'].split('\n')
sections = [l for l in sections if l]
callback(0.8, "Finish parsing.")
else:
raise NotImplementedError(
"file type not supported yet(doc, docx, pdf, txt supported)")
doc = {
"docnm_kwd": filename,
"title_tks": rag_tokenizer.tokenize(re.sub(r"\.[a-zA-Z]+$", "", filename))
}
doc["title_sm_tks"] = rag_tokenizer.fine_grained_tokenize(doc["title_tks"])
tokenize(doc, "\n".join(sections), eng)
return [doc]
if __name__ == "__main__":
import sys
def dummy(prog=None, msg=""):
pass
chunk(sys.argv[1], from_page=0, to_page=10, callback=dummy)

574
rag/app/paper.py
View File

@ -1,287 +1,287 @@
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
#
import copy
import re
from collections import Counter
from api.db import ParserType
from rag.nlp import rag_tokenizer, tokenize, tokenize_table, add_positions, bullets_category, title_frequency, tokenize_chunks
from deepdoc.parser import PdfParser, PlainParser
import numpy as np
from rag.utils import num_tokens_from_string
class Pdf(PdfParser):
def __init__(self):
self.model_speciess = ParserType.PAPER.value
super().__init__()
def __call__(self, filename, binary=None, from_page=0,
to_page=100000, zoomin=3, callback=None):
callback(msg="OCR is running...")
self.__images__(
filename if not binary else binary,
zoomin,
from_page,
to_page,
callback
)
callback(msg="OCR finished.")
from timeit import default_timer as timer
start = timer()
self._layouts_rec(zoomin)
callback(0.63, "Layout analysis finished")
print("layouts:", timer() - start)
self._table_transformer_job(zoomin)
callback(0.68, "Table analysis finished")
self._text_merge()
tbls = self._extract_table_figure(True, zoomin, True, True)
column_width = np.median([b["x1"] - b["x0"] for b in self.boxes])
self._concat_downward()
self._filter_forpages()
callback(0.75, "Text merging finished.")
# clean mess
if column_width < self.page_images[0].size[0] / zoomin / 2:
print("two_column...................", column_width,
self.page_images[0].size[0] / zoomin / 2)
self.boxes = self.sort_X_by_page(self.boxes, column_width / 2)
for b in self.boxes:
b["text"] = re.sub(r"([\t 　]|\u3000){2,}", " ", b["text"].strip())
def _begin(txt):
return re.match(
"[0-9. 一、i]*(introduction|abstract|摘要|引言|keywords|key words|关键词|background|背景|目录|前言|contents)",
txt.lower().strip())
if from_page > 0:
return {
"title": "",
"authors": "",
"abstract": "",
"sections": [(b["text"] + self._line_tag(b, zoomin), b.get("layoutno", "")) for b in self.boxes if
re.match(r"(text|title)", b.get("layoutno", "text"))],
"tables": tbls
}
# get title and authors
title = ""
authors = []
i = 0
while i < min(32, len(self.boxes)-1):
b = self.boxes[i]
i += 1
if b.get("layoutno", "").find("title") >= 0:
title = b["text"]
if _begin(title):
title = ""
break
for j in range(3):
if _begin(self.boxes[i + j]["text"]):
break
authors.append(self.boxes[i + j]["text"])
break
break
# get abstract
abstr = ""
i = 0
while i + 1 < min(32, len(self.boxes)):
b = self.boxes[i]
i += 1
txt = b["text"].lower().strip()
if re.match("(abstract|摘要)", txt):
if len(txt.split(" ")) > 32 or len(txt) > 64:
abstr = txt + self._line_tag(b, zoomin)
break
txt = self.boxes[i]["text"].lower().strip()
if len(txt.split(" ")) > 32 or len(txt) > 64:
abstr = txt + self._line_tag(self.boxes[i], zoomin)
i += 1
break
if not abstr:
i = 0
callback(
0.8, "Page {}~{}: Text merging finished".format(
from_page, min(
to_page, self.total_page)))
for b in self.boxes:
print(b["text"], b.get("layoutno"))
print(tbls)
return {
"title": title,
"authors": " ".join(authors),
"abstract": abstr,
"sections": [(b["text"] + self._line_tag(b, zoomin), b.get("layoutno", "")) for b in self.boxes[i:] if
re.match(r"(text|title)", b.get("layoutno", "text"))],
"tables": tbls
}
def chunk(filename, binary=None, from_page=0, to_page=100000,
lang="Chinese", callback=None, **kwargs):
"""
Only pdf is supported.
The abstract of the paper will be sliced as an entire chunk, and will not be sliced partly.
"""
pdf_parser = None
if re.search(r"\.pdf$", filename, re.IGNORECASE):
if not kwargs.get("parser_config", {}).get("layout_recognize", True):
pdf_parser = PlainParser()
paper = {
"title": filename,
"authors": " ",
"abstract": "",
"sections": pdf_parser(filename if not binary else binary, from_page=from_page, to_page=to_page)[0],
"tables": []
}
else:
pdf_parser = Pdf()
paper = pdf_parser(filename if not binary else binary,
from_page=from_page, to_page=to_page, callback=callback)
else:
raise NotImplementedError("file type not supported yet(pdf supported)")
doc = {"docnm_kwd": filename, "authors_tks": rag_tokenizer.tokenize(paper["authors"]),
"title_tks": rag_tokenizer.tokenize(paper["title"] if paper["title"] else filename)}
doc["title_sm_tks"] = rag_tokenizer.fine_grained_tokenize(doc["title_tks"])
doc["authors_sm_tks"] = rag_tokenizer.fine_grained_tokenize(doc["authors_tks"])
# is it English
eng = lang.lower() == "english" # pdf_parser.is_english
print("It's English.....", eng)
res = tokenize_table(paper["tables"], doc, eng)
if paper["abstract"]:
d = copy.deepcopy(doc)
txt = pdf_parser.remove_tag(paper["abstract"])
d["important_kwd"] = ["abstract", "总结", "概括", "summary", "summarize"]
d["important_tks"] = " ".join(d["important_kwd"])
d["image"], poss = pdf_parser.crop(
paper["abstract"], need_position=True)
add_positions(d, poss)
tokenize(d, txt, eng)
res.append(d)
sorted_sections = paper["sections"]
# set pivot using the most frequent type of title,
# then merge between 2 pivot
bull = bullets_category([txt for txt, _ in sorted_sections])
most_level, levels = title_frequency(bull, sorted_sections)
assert len(sorted_sections) == len(levels)
sec_ids = []
sid = 0
for i, lvl in enumerate(levels):
if lvl <= most_level and i > 0 and lvl != levels[i - 1]:
sid += 1
sec_ids.append(sid)
print(lvl, sorted_sections[i][0], most_level, sid)
chunks = []
last_sid = -2
for (txt, _), sec_id in zip(sorted_sections, sec_ids):
if sec_id == last_sid:
if chunks:
chunks[-1] += "\n" + txt
continue
chunks.append(txt)
last_sid = sec_id
res.extend(tokenize_chunks(chunks, doc, eng, pdf_parser))
return res
"""
readed = [0] * len(paper["lines"])
# find colon firstly
i = 0
while i + 1 < len(paper["lines"]):
txt = pdf_parser.remove_tag(paper["lines"][i][0])
j = i
if txt.strip("\n").strip()[-1] not in ":：":
i += 1
continue
i += 1
while i < len(paper["lines"]) and not paper["lines"][i][0]:
i += 1
if i >= len(paper["lines"]): break
proj = [paper["lines"][i][0].strip()]
i += 1
while i < len(paper["lines"]) and paper["lines"][i][0].strip()[0] == proj[-1][0]:
proj.append(paper["lines"][i])
i += 1
for k in range(j, i): readed[k] = True
txt = txt[::-1]
if eng:
r = re.search(r"(.*?) ([\\.;?!]|$)", txt)
txt = r.group(1)[::-1] if r else txt[::-1]
else:
r = re.search(r"(.*?) ([。？；！]|$)", txt)
txt = r.group(1)[::-1] if r else txt[::-1]
for p in proj:
d = copy.deepcopy(doc)
txt += "\n" + pdf_parser.remove_tag(p)
d["image"], poss = pdf_parser.crop(p, need_position=True)
add_positions(d, poss)
tokenize(d, txt, eng)
res.append(d)
i = 0
chunk = []
tk_cnt = 0
def add_chunk():
nonlocal chunk, res, doc, pdf_parser, tk_cnt
d = copy.deepcopy(doc)
ck = "\n".join(chunk)
tokenize(d, pdf_parser.remove_tag(ck), pdf_parser.is_english)
d["image"], poss = pdf_parser.crop(ck, need_position=True)
add_positions(d, poss)
res.append(d)
chunk = []
tk_cnt = 0
while i < len(paper["lines"]):
if tk_cnt > 128:
add_chunk()
if readed[i]:
i += 1
continue
readed[i] = True
txt, layouts = paper["lines"][i]
txt_ = pdf_parser.remove_tag(txt)
i += 1
cnt = num_tokens_from_string(txt_)
if any([
layouts.find("title") >= 0 and chunk,
cnt + tk_cnt > 128 and tk_cnt > 32,
]):
add_chunk()
chunk = [txt]
tk_cnt = cnt
else:
chunk.append(txt)
tk_cnt += cnt
if chunk: add_chunk()
for i, d in enumerate(res):
print(d)
# d["image"].save(f"./logs/{i}.jpg")
return res
"""
if __name__ == "__main__":
import sys
def dummy(prog=None, msg=""):
pass
chunk(sys.argv[1], callback=dummy)
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
#
import copy
import re
from collections import Counter
from api.db import ParserType
from rag.nlp import rag_tokenizer, tokenize, tokenize_table, add_positions, bullets_category, title_frequency, tokenize_chunks
from deepdoc.parser import PdfParser, PlainParser
import numpy as np
from rag.utils import num_tokens_from_string
class Pdf(PdfParser):
def __init__(self):
self.model_speciess = ParserType.PAPER.value
super().__init__()
def __call__(self, filename, binary=None, from_page=0,
to_page=100000, zoomin=3, callback=None):
callback(msg="OCR is running...")
self.__images__(
filename if not binary else binary,
zoomin,
from_page,
to_page,
callback
)
callback(msg="OCR finished.")
from timeit import default_timer as timer
start = timer()
self._layouts_rec(zoomin)
callback(0.63, "Layout analysis finished")
print("layouts:", timer() - start)
self._table_transformer_job(zoomin)
callback(0.68, "Table analysis finished")
self._text_merge()
tbls = self._extract_table_figure(True, zoomin, True, True)
column_width = np.median([b["x1"] - b["x0"] for b in self.boxes])
self._concat_downward()
self._filter_forpages()
callback(0.75, "Text merging finished.")
# clean mess
if column_width < self.page_images[0].size[0] / zoomin / 2:
print("two_column...................", column_width,
self.page_images[0].size[0] / zoomin / 2)
self.boxes = self.sort_X_by_page(self.boxes, column_width / 2)
for b in self.boxes:
b["text"] = re.sub(r"([\t 　]|\u3000){2,}", " ", b["text"].strip())
def _begin(txt):
return re.match(
"[0-9. 一、i]*(introduction|abstract|摘要|引言|keywords|key words|关键词|background|背景|目录|前言|contents)",
txt.lower().strip())
if from_page > 0:
return {
"title": "",
"authors": "",
"abstract": "",
"sections": [(b["text"] + self._line_tag(b, zoomin), b.get("layoutno", "")) for b in self.boxes if
re.match(r"(text|title)", b.get("layoutno", "text"))],
"tables": tbls
}
# get title and authors
title = ""
authors = []
i = 0
while i < min(32, len(self.boxes)-1):
b = self.boxes[i]
i += 1
if b.get("layoutno", "").find("title") >= 0:
title = b["text"]
if _begin(title):
title = ""
break
for j in range(3):
if _begin(self.boxes[i + j]["text"]):
break
authors.append(self.boxes[i + j]["text"])
break
break
# get abstract
abstr = ""
i = 0
while i + 1 < min(32, len(self.boxes)):
b = self.boxes[i]
i += 1
txt = b["text"].lower().strip()
if re.match("(abstract|摘要)", txt):
if len(txt.split(" ")) > 32 or len(txt) > 64:
abstr = txt + self._line_tag(b, zoomin)
break
txt = self.boxes[i]["text"].lower().strip()
if len(txt.split(" ")) > 32 or len(txt) > 64:
abstr = txt + self._line_tag(self.boxes[i], zoomin)
i += 1
break
if not abstr:
i = 0
callback(
0.8, "Page {}~{}: Text merging finished".format(
from_page, min(
to_page, self.total_page)))
for b in self.boxes:
print(b["text"], b.get("layoutno"))
print(tbls)
return {
"title": title,
"authors": " ".join(authors),
"abstract": abstr,
"sections": [(b["text"] + self._line_tag(b, zoomin), b.get("layoutno", "")) for b in self.boxes[i:] if
re.match(r"(text|title)", b.get("layoutno", "text"))],
"tables": tbls
}
def chunk(filename, binary=None, from_page=0, to_page=100000,
lang="Chinese", callback=None, **kwargs):
"""
Only pdf is supported.
The abstract of the paper will be sliced as an entire chunk, and will not be sliced partly.
"""
pdf_parser = None
if re.search(r"\.pdf$", filename, re.IGNORECASE):
if not kwargs.get("parser_config", {}).get("layout_recognize", True):
pdf_parser = PlainParser()
paper = {
"title": filename,
"authors": " ",
"abstract": "",
"sections": pdf_parser(filename if not binary else binary, from_page=from_page, to_page=to_page)[0],
"tables": []
}
else:
pdf_parser = Pdf()
paper = pdf_parser(filename if not binary else binary,
from_page=from_page, to_page=to_page, callback=callback)
else:
raise NotImplementedError("file type not supported yet(pdf supported)")
doc = {"docnm_kwd": filename, "authors_tks": rag_tokenizer.tokenize(paper["authors"]),
"title_tks": rag_tokenizer.tokenize(paper["title"] if paper["title"] else filename)}
doc["title_sm_tks"] = rag_tokenizer.fine_grained_tokenize(doc["title_tks"])
doc["authors_sm_tks"] = rag_tokenizer.fine_grained_tokenize(doc["authors_tks"])
# is it English
eng = lang.lower() == "english" # pdf_parser.is_english
print("It's English.....", eng)
res = tokenize_table(paper["tables"], doc, eng)
if paper["abstract"]:
d = copy.deepcopy(doc)
txt = pdf_parser.remove_tag(paper["abstract"])
d["important_kwd"] = ["abstract", "总结", "概括", "summary", "summarize"]
d["important_tks"] = " ".join(d["important_kwd"])
d["image"], poss = pdf_parser.crop(
paper["abstract"], need_position=True)
add_positions(d, poss)
tokenize(d, txt, eng)
res.append(d)
sorted_sections = paper["sections"]
# set pivot using the most frequent type of title,
# then merge between 2 pivot
bull = bullets_category([txt for txt, _ in sorted_sections])
most_level, levels = title_frequency(bull, sorted_sections)
assert len(sorted_sections) == len(levels)
sec_ids = []
sid = 0
for i, lvl in enumerate(levels):
if lvl <= most_level and i > 0 and lvl != levels[i - 1]:
sid += 1
sec_ids.append(sid)
print(lvl, sorted_sections[i][0], most_level, sid)
chunks = []
last_sid = -2
for (txt, _), sec_id in zip(sorted_sections, sec_ids):
if sec_id == last_sid:
if chunks:
chunks[-1] += "\n" + txt
continue
chunks.append(txt)
last_sid = sec_id
res.extend(tokenize_chunks(chunks, doc, eng, pdf_parser))
return res
"""
readed = [0] * len(paper["lines"])
# find colon firstly
i = 0
while i + 1 < len(paper["lines"]):
txt = pdf_parser.remove_tag(paper["lines"][i][0])
j = i
if txt.strip("\n").strip()[-1] not in ":：":
i += 1
continue
i += 1
while i < len(paper["lines"]) and not paper["lines"][i][0]:
i += 1
if i >= len(paper["lines"]): break
proj = [paper["lines"][i][0].strip()]
i += 1
while i < len(paper["lines"]) and paper["lines"][i][0].strip()[0] == proj[-1][0]:
proj.append(paper["lines"][i])
i += 1
for k in range(j, i): readed[k] = True
txt = txt[::-1]
if eng:
r = re.search(r"(.*?) ([\\.;?!]|$)", txt)
txt = r.group(1)[::-1] if r else txt[::-1]
else:
r = re.search(r"(.*?) ([。？；！]|$)", txt)
txt = r.group(1)[::-1] if r else txt[::-1]
for p in proj:
d = copy.deepcopy(doc)
txt += "\n" + pdf_parser.remove_tag(p)
d["image"], poss = pdf_parser.crop(p, need_position=True)
add_positions(d, poss)
tokenize(d, txt, eng)
res.append(d)
i = 0
chunk = []
tk_cnt = 0
def add_chunk():
nonlocal chunk, res, doc, pdf_parser, tk_cnt
d = copy.deepcopy(doc)
ck = "\n".join(chunk)
tokenize(d, pdf_parser.remove_tag(ck), pdf_parser.is_english)
d["image"], poss = pdf_parser.crop(ck, need_position=True)
add_positions(d, poss)
res.append(d)
chunk = []
tk_cnt = 0
while i < len(paper["lines"]):
if tk_cnt > 128:
add_chunk()
if readed[i]:
i += 1
continue
readed[i] = True
txt, layouts = paper["lines"][i]
txt_ = pdf_parser.remove_tag(txt)
i += 1
cnt = num_tokens_from_string(txt_)
if any([
layouts.find("title") >= 0 and chunk,
cnt + tk_cnt > 128 and tk_cnt > 32,
]):
add_chunk()
chunk = [txt]
tk_cnt = cnt
else:
chunk.append(txt)
tk_cnt += cnt
if chunk: add_chunk()
for i, d in enumerate(res):
print(d)
# d["image"].save(f"./logs/{i}.jpg")
return res
"""
if __name__ == "__main__":
import sys
def dummy(prog=None, msg=""):
pass
chunk(sys.argv[1], callback=dummy)

104
rag/app/picture.py
View File

@ -1,52 +1,52 @@
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
#
import io
import numpy as np
from PIL import Image
from api.db import LLMType
from api.db.services.llm_service import LLMBundle
from rag.nlp import tokenize
from deepdoc.vision import OCR
ocr = OCR()
def chunk(filename, binary, tenant_id, lang, callback=None, **kwargs):
img = Image.open(io.BytesIO(binary)).convert('RGB')
doc = {
"docnm_kwd": filename,
"image": img
}
bxs = ocr(np.array(img))
txt = "\n".join([t[0] for _, t in bxs if t[0]])
eng = lang.lower() == "english"
callback(0.4, "Finish OCR: (%s ...)" % txt[:12])
if (eng and len(txt.split(" ")) > 32) or len(txt) > 32:
tokenize(doc, txt, eng)
callback(0.8, "OCR results is too long to use CV LLM.")
return [doc]
try:
callback(0.4, "Use CV LLM to describe the picture.")
cv_mdl = LLMBundle(tenant_id, LLMType.IMAGE2TEXT, lang=lang)
ans = cv_mdl.describe(binary)
callback(0.8, "CV LLM respond: %s ..." % ans[:32])
txt += "\n" + ans
tokenize(doc, txt, eng)
return [doc]
except Exception as e:
callback(prog=-1, msg=str(e))
return []
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
#
import io
import numpy as np
from PIL import Image
from api.db import LLMType
from api.db.services.llm_service import LLMBundle
from rag.nlp import tokenize
from deepdoc.vision import OCR
ocr = OCR()
def chunk(filename, binary, tenant_id, lang, callback=None, **kwargs):
img = Image.open(io.BytesIO(binary)).convert('RGB')
doc = {
"docnm_kwd": filename,
"image": img
}
bxs = ocr(np.array(img))
txt = "\n".join([t[0] for _, t in bxs if t[0]])
eng = lang.lower() == "english"
callback(0.4, "Finish OCR: (%s ...)" % txt[:12])
if (eng and len(txt.split(" ")) > 32) or len(txt) > 32:
tokenize(doc, txt, eng)
callback(0.8, "OCR results is too long to use CV LLM.")
return [doc]
try:
callback(0.4, "Use CV LLM to describe the picture.")
cv_mdl = LLMBundle(tenant_id, LLMType.IMAGE2TEXT, lang=lang)
ans = cv_mdl.describe(binary)
callback(0.8, "CV LLM respond: %s ..." % ans[:32])
txt += "\n" + ans
tokenize(doc, txt, eng)
return [doc]
except Exception as e:
callback(prog=-1, msg=str(e))
return []

286
rag/app/presentation.py
View File

@ -1,143 +1,143 @@
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
#
import copy
import re
from io import BytesIO
from PIL import Image
from rag.nlp import tokenize, is_english
from rag.nlp import rag_tokenizer
from deepdoc.parser import PdfParser, PptParser, PlainParser
from PyPDF2 import PdfReader as pdf2_read
class Ppt(PptParser):
def __call__(self, fnm, from_page, to_page, callback=None):
txts = super().__call__(fnm, from_page, to_page)
callback(0.5, "Text extraction finished.")
import aspose.slides as slides
import aspose.pydrawing as drawing
imgs = []
with slides.Presentation(BytesIO(fnm)) as presentation:
for i, slide in enumerate(presentation.slides[from_page: to_page]):
buffered = BytesIO()
slide.get_thumbnail(
0.5, 0.5).save(
buffered, drawing.imaging.ImageFormat.jpeg)
imgs.append(Image.open(buffered))
assert len(imgs) == len(
txts), "Slides text and image do not match: {} vs. {}".format(len(imgs), len(txts))
callback(0.9, "Image extraction finished")
self.is_english = is_english(txts)
return [(txts[i], imgs[i]) for i in range(len(txts))]
class Pdf(PdfParser):
def __init__(self):
super().__init__()
def __garbage(self, txt):
txt = txt.lower().strip()
if re.match(r"[0-9\.,%/-]+$", txt):
return True
if len(txt) < 3:
return True
return False
def __call__(self, filename, binary=None, from_page=0,
to_page=100000, zoomin=3, callback=None):
callback(msg="OCR is running...")
self.__images__(filename if not binary else binary,
zoomin, from_page, to_page, callback)
callback(0.8, "Page {}~{}: OCR finished".format(
from_page, min(to_page, self.total_page)))
assert len(self.boxes) == len(self.page_images), "{} vs. {}".format(
len(self.boxes), len(self.page_images))
res = []
for i in range(len(self.boxes)):
lines = "\n".join([b["text"] for b in self.boxes[i]
if not self.__garbage(b["text"])])
res.append((lines, self.page_images[i]))
callback(0.9, "Page {}~{}: Parsing finished".format(
from_page, min(to_page, self.total_page)))
return res
class PlainPdf(PlainParser):
def __call__(self, filename, binary=None, from_page=0,
to_page=100000, callback=None, **kwargs):
self.pdf = pdf2_read(filename if not binary else BytesIO(binary))
page_txt = []
for page in self.pdf.pages[from_page: to_page]:
page_txt.append(page.extract_text())
callback(0.9, "Parsing finished")
return [(txt, None) for txt in page_txt]
def chunk(filename, binary=None, from_page=0, to_page=100000,
lang="Chinese", callback=None, **kwargs):
"""
The supported file formats are pdf, pptx.
Every page will be treated as a chunk. And the thumbnail of every page will be stored.
PPT file will be parsed by using this method automatically, setting-up for every PPT file is not necessary.
"""
eng = lang.lower() == "english"
doc = {
"docnm_kwd": filename,
"title_tks": rag_tokenizer.tokenize(re.sub(r"\.[a-zA-Z]+$", "", filename))
}
doc["title_sm_tks"] = rag_tokenizer.fine_grained_tokenize(doc["title_tks"])
res = []
if re.search(r"\.pptx?$", filename, re.IGNORECASE):
ppt_parser = Ppt()
for pn, (txt, img) in enumerate(ppt_parser(
filename if not binary else binary, from_page, 1000000, callback)):
d = copy.deepcopy(doc)
pn += from_page
d["image"] = img
d["page_num_int"] = [pn + 1]
d["top_int"] = [0]
d["position_int"] = [(pn + 1, 0, img.size[0], 0, img.size[1])]
tokenize(d, txt, eng)
res.append(d)
return res
elif re.search(r"\.pdf$", filename, re.IGNORECASE):
pdf_parser = Pdf() if kwargs.get(
"parser_config", {}).get(
"layout_recognize", True) else PlainPdf()
for pn, (txt, img) in enumerate(pdf_parser(filename, binary,
from_page=from_page, to_page=to_page, callback=callback)):
d = copy.deepcopy(doc)
pn += from_page
if img:
d["image"] = img
d["page_num_int"] = [pn + 1]
d["top_int"] = [0]
d["position_int"] = [
(pn + 1, 0, img.size[0] if img else 0, 0, img.size[1] if img else 0)]
tokenize(d, txt, eng)
res.append(d)
return res
raise NotImplementedError(
"file type not supported yet(pptx, pdf supported)")
if __name__ == "__main__":
import sys
def dummy(a, b):
pass
chunk(sys.argv[1], callback=dummy)
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
#
import copy
import re
from io import BytesIO
from PIL import Image
from rag.nlp import tokenize, is_english
from rag.nlp import rag_tokenizer
from deepdoc.parser import PdfParser, PptParser, PlainParser
from PyPDF2 import PdfReader as pdf2_read
class Ppt(PptParser):
def __call__(self, fnm, from_page, to_page, callback=None):
txts = super().__call__(fnm, from_page, to_page)
callback(0.5, "Text extraction finished.")
import aspose.slides as slides
import aspose.pydrawing as drawing
imgs = []
with slides.Presentation(BytesIO(fnm)) as presentation:
for i, slide in enumerate(presentation.slides[from_page: to_page]):
buffered = BytesIO()
slide.get_thumbnail(
0.5, 0.5).save(
buffered, drawing.imaging.ImageFormat.jpeg)
imgs.append(Image.open(buffered))
assert len(imgs) == len(
txts), "Slides text and image do not match: {} vs. {}".format(len(imgs), len(txts))
callback(0.9, "Image extraction finished")
self.is_english = is_english(txts)
return [(txts[i], imgs[i]) for i in range(len(txts))]
class Pdf(PdfParser):
def __init__(self):
super().__init__()
def __garbage(self, txt):
txt = txt.lower().strip()
if re.match(r"[0-9\.,%/-]+$", txt):
return True
if len(txt) < 3:
return True
return False
def __call__(self, filename, binary=None, from_page=0,
to_page=100000, zoomin=3, callback=None):
callback(msg="OCR is running...")
self.__images__(filename if not binary else binary,
zoomin, from_page, to_page, callback)
callback(0.8, "Page {}~{}: OCR finished".format(
from_page, min(to_page, self.total_page)))
assert len(self.boxes) == len(self.page_images), "{} vs. {}".format(
len(self.boxes), len(self.page_images))
res = []
for i in range(len(self.boxes)):
lines = "\n".join([b["text"] for b in self.boxes[i]
if not self.__garbage(b["text"])])
res.append((lines, self.page_images[i]))
callback(0.9, "Page {}~{}: Parsing finished".format(
from_page, min(to_page, self.total_page)))
return res
class PlainPdf(PlainParser):
def __call__(self, filename, binary=None, from_page=0,
to_page=100000, callback=None, **kwargs):
self.pdf = pdf2_read(filename if not binary else BytesIO(binary))
page_txt = []
for page in self.pdf.pages[from_page: to_page]:
page_txt.append(page.extract_text())
callback(0.9, "Parsing finished")
return [(txt, None) for txt in page_txt]
def chunk(filename, binary=None, from_page=0, to_page=100000,
lang="Chinese", callback=None, **kwargs):
"""
The supported file formats are pdf, pptx.
Every page will be treated as a chunk. And the thumbnail of every page will be stored.
PPT file will be parsed by using this method automatically, setting-up for every PPT file is not necessary.
"""
eng = lang.lower() == "english"
doc = {
"docnm_kwd": filename,
"title_tks": rag_tokenizer.tokenize(re.sub(r"\.[a-zA-Z]+$", "", filename))
}
doc["title_sm_tks"] = rag_tokenizer.fine_grained_tokenize(doc["title_tks"])
res = []
if re.search(r"\.pptx?$", filename, re.IGNORECASE):
ppt_parser = Ppt()
for pn, (txt, img) in enumerate(ppt_parser(
filename if not binary else binary, from_page, 1000000, callback)):
d = copy.deepcopy(doc)
pn += from_page
d["image"] = img
d["page_num_int"] = [pn + 1]
d["top_int"] = [0]
d["position_int"] = [(pn + 1, 0, img.size[0], 0, img.size[1])]
tokenize(d, txt, eng)
res.append(d)
return res
elif re.search(r"\.pdf$", filename, re.IGNORECASE):
pdf_parser = Pdf() if kwargs.get(
"parser_config", {}).get(
"layout_recognize", True) else PlainPdf()
for pn, (txt, img) in enumerate(pdf_parser(filename, binary,
from_page=from_page, to_page=to_page, callback=callback)):
d = copy.deepcopy(doc)
pn += from_page
if img:
d["image"] = img
d["page_num_int"] = [pn + 1]
d["top_int"] = [0]
d["position_int"] = [
(pn + 1, 0, img.size[0] if img else 0, 0, img.size[1] if img else 0)]
tokenize(d, txt, eng)
res.append(d)
return res
raise NotImplementedError(
"file type not supported yet(pptx, pdf supported)")
if __name__ == "__main__":
import sys
def dummy(a, b):
pass
chunk(sys.argv[1], callback=dummy)

842
rag/app/qa.py
View File

@ -1,422 +1,422 @@
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
#
import re
from copy import deepcopy
from io import BytesIO
from timeit import default_timer as timer
from nltk import word_tokenize
from openpyxl import load_workbook
from rag.nlp import is_english, random_choices, find_codec, qbullets_category, add_positions, has_qbullet, docx_question_level
from rag.nlp import rag_tokenizer, tokenize_table, concat_img
from rag.settings import cron_logger
from deepdoc.parser import PdfParser, ExcelParser, DocxParser
from docx import Document
from PIL import Image
from markdown import markdown
class Excel(ExcelParser):
def __call__(self, fnm, binary=None, callback=None):
if not binary:
wb = load_workbook(fnm)
else:
wb = load_workbook(BytesIO(binary))
total = 0
for sheetname in wb.sheetnames:
total += len(list(wb[sheetname].rows))
res, fails = [], []
for sheetname in wb.sheetnames:
ws = wb[sheetname]
rows = list(ws.rows)
for i, r in enumerate(rows):
q, a = "", ""
for cell in r:
if not cell.value:
continue
if not q:
q = str(cell.value)
elif not a:
a = str(cell.value)
else:
break
if q and a:
res.append((q, a))
else:
fails.append(str(i + 1))
if len(res) % 999 == 0:
callback(len(res) *
0.6 /
total, ("Extract Q&A: {}".format(len(res)) +
(f"{len(fails)} failure, line: %s..." %
(",".join(fails[:3])) if fails else "")))
callback(0.6, ("Extract Q&A: {}. ".format(len(res)) + (
f"{len(fails)} failure, line: %s..." % (",".join(fails[:3])) if fails else "")))
self.is_english = is_english(
[rmPrefix(q) for q, _ in random_choices(res, k=30) if len(q) > 1])
return res
class Pdf(PdfParser):
def __call__(self, filename, binary=None, from_page=0,
to_page=100000, zoomin=3, callback=None):
start = timer()
callback(msg="OCR is running...")
self.__images__(
filename if not binary else binary,
zoomin,
from_page,
to_page,
callback
)
callback(msg="OCR finished")
cron_logger.info("OCR({}~{}): {}".format(from_page, to_page, timer() - start))
start = timer()
self._layouts_rec(zoomin, drop=False)
callback(0.63, "Layout analysis finished.")
self._table_transformer_job(zoomin)
callback(0.65, "Table analysis finished.")
self._text_merge()
callback(0.67, "Text merging finished")
tbls = self._extract_table_figure(True, zoomin, True, True)
#self._naive_vertical_merge()
# self._concat_downward()
#self._filter_forpages()
cron_logger.info("layouts: {}".format(timer() - start))
sections = [b["text"] for b in self.boxes]
bull_x0_list = []
q_bull, reg = qbullets_category(sections)
if q_bull == -1:
raise ValueError("Unable to recognize Q&A structure.")
qai_list = []
last_q, last_a, last_tag = '', '', ''
last_index = -1
last_box = {'text':''}
last_bull = None
def sort_key(element):
tbls_pn = element[1][0][0]
tbls_top = element[1][0][3]
return tbls_pn, tbls_top
tbls.sort(key=sort_key)
tbl_index = 0
last_pn, last_bottom = 0, 0
tbl_pn, tbl_left, tbl_right, tbl_top, tbl_bottom, tbl_tag, tbl_text = 1, 0, 0, 0, 0, '@@0\t0\t0\t0\t0##', ''
for box in self.boxes:
section, line_tag = box['text'], self._line_tag(box, zoomin)
has_bull, index = has_qbullet(reg, box, last_box, last_index, last_bull, bull_x0_list)
last_box, last_index, last_bull = box, index, has_bull
line_pn = float(line_tag.lstrip('@@').split('\t')[0])
line_top = float(line_tag.rstrip('##').split('\t')[3])
tbl_pn, tbl_left, tbl_right, tbl_top, tbl_bottom, tbl_tag, tbl_text = self.get_tbls_info(tbls, tbl_index)
if not has_bull: # No question bullet
if not last_q:
if tbl_pn < line_pn or (tbl_pn == line_pn and tbl_top <= line_top): # image passed
tbl_index += 1
continue
else:
sum_tag = line_tag
sum_section = section
while ((tbl_pn == last_pn and tbl_top>= last_bottom) or (tbl_pn > last_pn)) \
and ((tbl_pn == line_pn and tbl_top <= line_top) or (tbl_pn < line_pn)): # add image at the middle of current answer
sum_tag = f'{tbl_tag}{sum_tag}'
sum_section = f'{tbl_text}{sum_section}'
tbl_index += 1
tbl_pn, tbl_left, tbl_right, tbl_top, tbl_bottom, tbl_tag, tbl_text = self.get_tbls_info(tbls, tbl_index)
last_a = f'{last_a}{sum_section}'
last_tag = f'{last_tag}{sum_tag}'
else:
if last_q:
while ((tbl_pn == last_pn and tbl_top>= last_bottom) or (tbl_pn > last_pn)) \
and ((tbl_pn == line_pn and tbl_top <= line_top) or (tbl_pn < line_pn)): # add image at the end of last answer
last_tag = f'{last_tag}{tbl_tag}'
last_a = f'{last_a}{tbl_text}'
tbl_index += 1
tbl_pn, tbl_left, tbl_right, tbl_top, tbl_bottom, tbl_tag, tbl_text = self.get_tbls_info(tbls, tbl_index)
image, poss = self.crop(last_tag, need_position=True)
qai_list.append((last_q, last_a, image, poss))
last_q, last_a, last_tag = '', '', ''
last_q = has_bull.group()
_, end = has_bull.span()
last_a = section[end:]
last_tag = line_tag
last_bottom = float(line_tag.rstrip('##').split('\t')[4])
last_pn = line_pn
if last_q:
qai_list.append((last_q, last_a, *self.crop(last_tag, need_position=True)))
return qai_list, tbls
def get_tbls_info(self, tbls, tbl_index):
if tbl_index >= len(tbls):
return 1, 0, 0, 0, 0, '@@0\t0\t0\t0\t0##', ''
tbl_pn = tbls[tbl_index][1][0][0]+1
tbl_left = tbls[tbl_index][1][0][1]
tbl_right = tbls[tbl_index][1][0][2]
tbl_top = tbls[tbl_index][1][0][3]
tbl_bottom = tbls[tbl_index][1][0][4]
tbl_tag = "@@{}\t{:.1f}\t{:.1f}\t{:.1f}\t{:.1f}##" \
.format(tbl_pn, tbl_left, tbl_right, tbl_top, tbl_bottom)
tbl_text = ''.join(tbls[tbl_index][0][1])
return tbl_pn, tbl_left, tbl_right, tbl_top, tbl_bottom, tbl_tag, tbl_text
class Docx(DocxParser):
def __init__(self):
pass
def get_picture(self, document, paragraph):
img = paragraph._element.xpath('.//pic:pic')
if not img:
return None
img = img[0]
embed = img.xpath('.//a:blip/@r:embed')[0]
related_part = document.part.related_parts[embed]
image = related_part.image
image = Image.open(BytesIO(image.blob)).convert('RGB')
return image
def __call__(self, filename, binary=None, from_page=0, to_page=100000, callback=None):
self.doc = Document(
filename) if not binary else Document(BytesIO(binary))
pn = 0
last_answer, last_image = "", None
question_stack, level_stack = [], []
qai_list = []
for p in self.doc.paragraphs:
if pn > to_page:
break
question_level, p_text = 0, ''
if from_page <= pn < to_page and p.text.strip():
question_level, p_text = docx_question_level(p)
if not question_level or question_level > 6: # not a question
last_answer = f'{last_answer}\n{p_text}'
current_image = self.get_picture(self.doc, p)
last_image = concat_img(last_image, current_image)
else: # is a question
if last_answer or last_image:
sum_question = '\n'.join(question_stack)
if sum_question:
qai_list.append((sum_question, last_answer, last_image))
last_answer, last_image = '', None
i = question_level
while question_stack and i <= level_stack[-1]:
question_stack.pop()
level_stack.pop()
question_stack.append(p_text)
level_stack.append(question_level)
for run in p.runs:
if 'lastRenderedPageBreak' in run._element.xml:
pn += 1
continue
if 'w:br' in run._element.xml and 'type="page"' in run._element.xml:
pn += 1
if last_answer:
sum_question = '\n'.join(question_stack)
if sum_question:
qai_list.append((sum_question, last_answer, last_image))
tbls = []
for tb in self.doc.tables:
html= "<table>"
for r in tb.rows:
html += "<tr>"
i = 0
while i < len(r.cells):
span = 1
c = r.cells[i]
for j in range(i+1, len(r.cells)):
if c.text == r.cells[j].text:
span += 1
i = j
i += 1
html += f"<td>{c.text}</td>" if span == 1 else f"<td colspan='{span}'>{c.text}</td>"
html += "</tr>"
html += "</table>"
tbls.append(((None, html), ""))
return qai_list, tbls
def rmPrefix(txt):
return re.sub(
r"^(问题|答案|回答|user|assistant|Q|A|Question|Answer|问|答)[\t:： ]+", "", txt.strip(), flags=re.IGNORECASE)
def beAdocPdf(d, q, a, eng, image, poss):
qprefix = "Question: " if eng else "问题："
aprefix = "Answer: " if eng else "回答："
d["content_with_weight"] = "\t".join(
[qprefix + rmPrefix(q), aprefix + rmPrefix(a)])
d["content_ltks"] = rag_tokenizer.tokenize(q)
d["content_sm_ltks"] = rag_tokenizer.fine_grained_tokenize(d["content_ltks"])
d["image"] = image
add_positions(d, poss)
return d
def beAdocDocx(d, q, a, eng, image):
qprefix = "Question: " if eng else "问题："
aprefix = "Answer: " if eng else "回答："
d["content_with_weight"] = "\t".join(
[qprefix + rmPrefix(q), aprefix + rmPrefix(a)])
d["content_ltks"] = rag_tokenizer.tokenize(q)
d["content_sm_ltks"] = rag_tokenizer.fine_grained_tokenize(d["content_ltks"])
d["image"] = image
return d
def beAdoc(d, q, a, eng):
qprefix = "Question: " if eng else "问题："
aprefix = "Answer: " if eng else "回答："
d["content_with_weight"] = "\t".join(
[qprefix + rmPrefix(q), aprefix + rmPrefix(a)])
d["content_ltks"] = rag_tokenizer.tokenize(q)
d["content_sm_ltks"] = rag_tokenizer.fine_grained_tokenize(d["content_ltks"])
return d
def mdQuestionLevel(s):
match = re.match(r'#*', s)
return (len(match.group(0)), s.lstrip('#').lstrip()) if match else (0, s)
def chunk(filename, binary=None, lang="Chinese", callback=None, **kwargs):
"""
Excel and csv(txt) format files are supported.
If the file is in excel format, there should be 2 column question and answer without header.
And question column is ahead of answer column.
And it's O.K if it has multiple sheets as long as the columns are rightly composed.
If it's in csv format, it should be UTF-8 encoded. Use TAB as delimiter to separate question and answer.
All the deformed lines will be ignored.
Every pair of Q&A will be treated as a chunk.
"""
eng = lang.lower() == "english"
res = []
doc = {
"docnm_kwd": filename,
"title_tks": rag_tokenizer.tokenize(re.sub(r"\.[a-zA-Z]+$", "", filename))
}
if re.search(r"\.xlsx?$", filename, re.IGNORECASE):
callback(0.1, "Start to parse.")
excel_parser = Excel()
for q, a in excel_parser(filename, binary, callback):
res.append(beAdoc(deepcopy(doc), q, a, eng))
return res
elif re.search(r"\.(txt|csv)$", filename, re.IGNORECASE):
callback(0.1, "Start to parse.")
txt = ""
if binary:
encoding = find_codec(binary)
txt = binary.decode(encoding, errors="ignore")
else:
with open(filename, "r") as f:
while True:
l = f.readline()
if not l:
break
txt += l
lines = txt.split("\n")
comma, tab = 0, 0
for l in lines:
if len(l.split(",")) == 2: comma += 1
if len(l.split("\t")) == 2: tab += 1
delimiter = "\t" if tab >= comma else ","
fails = []
question, answer = "", ""
i = 0
while i < len(lines):
arr = lines[i].split(delimiter)
if len(arr) != 2:
if question: answer += "\n" + lines[i]
else:
fails.append(str(i+1))
elif len(arr) == 2:
if question and answer: res.append(beAdoc(deepcopy(doc), question, answer, eng))
question, answer = arr
i += 1
if len(res) % 999 == 0:
callback(len(res) * 0.6 / len(lines), ("Extract Q&A: {}".format(len(res)) + (
f"{len(fails)} failure, line: %s..." % (",".join(fails[:3])) if fails else "")))
if question: res.append(beAdoc(deepcopy(doc), question, answer, eng))
callback(0.6, ("Extract Q&A: {}".format(len(res)) + (
f"{len(fails)} failure, line: %s..." % (",".join(fails[:3])) if fails else "")))
return res
elif re.search(r"\.pdf$", filename, re.IGNORECASE):
callback(0.1, "Start to parse.")
pdf_parser = Pdf()
qai_list, tbls = pdf_parser(filename if not binary else binary,
from_page=0, to_page=10000, callback=callback)
for q, a, image, poss in qai_list:
res.append(beAdocPdf(deepcopy(doc), q, a, eng, image, poss))
return res
elif re.search(r"\.(md|markdown)$", filename, re.IGNORECASE):
callback(0.1, "Start to parse.")
txt = ""
if binary:
encoding = find_codec(binary)
txt = binary.decode(encoding, errors="ignore")
else:
with open(filename, "r") as f:
while True:
l = f.readline()
if not l:
break
txt += l
lines = txt.split("\n")
last_question, last_answer = "", ""
question_stack, level_stack = [], []
code_block = False
level_index = [-1] * 7
for index, l in enumerate(lines):
if l.strip().startswith('```'):
code_block = not code_block
question_level, question = 0, ''
if not code_block:
question_level, question = mdQuestionLevel(l)
if not question_level or question_level > 6: # not a question
last_answer = f'{last_answer}\n{l}'
else: # is a question
if last_answer.strip():
sum_question = '\n'.join(question_stack)
if sum_question:
res.append(beAdoc(deepcopy(doc), sum_question, markdown(last_answer, extensions=['markdown.extensions.tables']), eng))
last_answer = ''
i = question_level
while question_stack and i <= level_stack[-1]:
question_stack.pop()
level_stack.pop()
question_stack.append(question)
level_stack.append(question_level)
if last_answer.strip():
sum_question = '\n'.join(question_stack)
if sum_question:
res.append(beAdoc(deepcopy(doc), sum_question, markdown(last_answer, extensions=['markdown.extensions.tables']), eng))
return res
elif re.search(r"\.docx$", filename, re.IGNORECASE):
docx_parser = Docx()
qai_list, tbls = docx_parser(filename, binary,
from_page=0, to_page=10000, callback=callback)
res = tokenize_table(tbls, doc, eng)
for q, a, image in qai_list:
res.append(beAdocDocx(deepcopy(doc), q, a, eng, image))
return res
raise NotImplementedError(
"Excel, csv(txt), pdf, markdown and docx format files are supported.")
if __name__ == "__main__":
import sys
def dummy(prog=None, msg=""):
pass
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
#
import re
from copy import deepcopy
from io import BytesIO
from timeit import default_timer as timer
from nltk import word_tokenize
from openpyxl import load_workbook
from rag.nlp import is_english, random_choices, find_codec, qbullets_category, add_positions, has_qbullet, docx_question_level
from rag.nlp import rag_tokenizer, tokenize_table, concat_img
from rag.settings import cron_logger
from deepdoc.parser import PdfParser, ExcelParser, DocxParser
from docx import Document
from PIL import Image
from markdown import markdown
class Excel(ExcelParser):
def __call__(self, fnm, binary=None, callback=None):
if not binary:
wb = load_workbook(fnm)
else:
wb = load_workbook(BytesIO(binary))
total = 0
for sheetname in wb.sheetnames:
total += len(list(wb[sheetname].rows))
res, fails = [], []
for sheetname in wb.sheetnames:
ws = wb[sheetname]
rows = list(ws.rows)
for i, r in enumerate(rows):
q, a = "", ""
for cell in r:
if not cell.value:
continue
if not q:
q = str(cell.value)
elif not a:
a = str(cell.value)
else:
break
if q and a:
res.append((q, a))
else:
fails.append(str(i + 1))
if len(res) % 999 == 0:
callback(len(res) *
0.6 /
total, ("Extract Q&A: {}".format(len(res)) +
(f"{len(fails)} failure, line: %s..." %
(",".join(fails[:3])) if fails else "")))
callback(0.6, ("Extract Q&A: {}. ".format(len(res)) + (
f"{len(fails)} failure, line: %s..." % (",".join(fails[:3])) if fails else "")))
self.is_english = is_english(
[rmPrefix(q) for q, _ in random_choices(res, k=30) if len(q) > 1])
return res
class Pdf(PdfParser):
def __call__(self, filename, binary=None, from_page=0,
to_page=100000, zoomin=3, callback=None):
start = timer()
callback(msg="OCR is running...")
self.__images__(
filename if not binary else binary,
zoomin,
from_page,
to_page,
callback
)
callback(msg="OCR finished")
cron_logger.info("OCR({}~{}): {}".format(from_page, to_page, timer() - start))
start = timer()
self._layouts_rec(zoomin, drop=False)
callback(0.63, "Layout analysis finished.")
self._table_transformer_job(zoomin)
callback(0.65, "Table analysis finished.")
self._text_merge()
callback(0.67, "Text merging finished")
tbls = self._extract_table_figure(True, zoomin, True, True)
#self._naive_vertical_merge()
# self._concat_downward()
#self._filter_forpages()
cron_logger.info("layouts: {}".format(timer() - start))
sections = [b["text"] for b in self.boxes]
bull_x0_list = []
q_bull, reg = qbullets_category(sections)
if q_bull == -1:
raise ValueError("Unable to recognize Q&A structure.")
qai_list = []
last_q, last_a, last_tag = '', '', ''
last_index = -1
last_box = {'text':''}
last_bull = None
def sort_key(element):
tbls_pn = element[1][0][0]
tbls_top = element[1][0][3]
return tbls_pn, tbls_top
tbls.sort(key=sort_key)
tbl_index = 0
last_pn, last_bottom = 0, 0
tbl_pn, tbl_left, tbl_right, tbl_top, tbl_bottom, tbl_tag, tbl_text = 1, 0, 0, 0, 0, '@@0\t0\t0\t0\t0##', ''
for box in self.boxes:
section, line_tag = box['text'], self._line_tag(box, zoomin)
has_bull, index = has_qbullet(reg, box, last_box, last_index, last_bull, bull_x0_list)
last_box, last_index, last_bull = box, index, has_bull
line_pn = float(line_tag.lstrip('@@').split('\t')[0])
line_top = float(line_tag.rstrip('##').split('\t')[3])
tbl_pn, tbl_left, tbl_right, tbl_top, tbl_bottom, tbl_tag, tbl_text = self.get_tbls_info(tbls, tbl_index)
if not has_bull: # No question bullet
if not last_q:
if tbl_pn < line_pn or (tbl_pn == line_pn and tbl_top <= line_top): # image passed
tbl_index += 1
continue
else:
sum_tag = line_tag
sum_section = section
while ((tbl_pn == last_pn and tbl_top>= last_bottom) or (tbl_pn > last_pn)) \
and ((tbl_pn == line_pn and tbl_top <= line_top) or (tbl_pn < line_pn)): # add image at the middle of current answer
sum_tag = f'{tbl_tag}{sum_tag}'
sum_section = f'{tbl_text}{sum_section}'
tbl_index += 1
tbl_pn, tbl_left, tbl_right, tbl_top, tbl_bottom, tbl_tag, tbl_text = self.get_tbls_info(tbls, tbl_index)
last_a = f'{last_a}{sum_section}'
last_tag = f'{last_tag}{sum_tag}'
else:
if last_q:
while ((tbl_pn == last_pn and tbl_top>= last_bottom) or (tbl_pn > last_pn)) \
and ((tbl_pn == line_pn and tbl_top <= line_top) or (tbl_pn < line_pn)): # add image at the end of last answer
last_tag = f'{last_tag}{tbl_tag}'
last_a = f'{last_a}{tbl_text}'
tbl_index += 1
tbl_pn, tbl_left, tbl_right, tbl_top, tbl_bottom, tbl_tag, tbl_text = self.get_tbls_info(tbls, tbl_index)
image, poss = self.crop(last_tag, need_position=True)
qai_list.append((last_q, last_a, image, poss))
last_q, last_a, last_tag = '', '', ''
last_q = has_bull.group()
_, end = has_bull.span()
last_a = section[end:]
last_tag = line_tag
last_bottom = float(line_tag.rstrip('##').split('\t')[4])
last_pn = line_pn
if last_q:
qai_list.append((last_q, last_a, *self.crop(last_tag, need_position=True)))
return qai_list, tbls
def get_tbls_info(self, tbls, tbl_index):
if tbl_index >= len(tbls):
return 1, 0, 0, 0, 0, '@@0\t0\t0\t0\t0##', ''
tbl_pn = tbls[tbl_index][1][0][0]+1
tbl_left = tbls[tbl_index][1][0][1]
tbl_right = tbls[tbl_index][1][0][2]
tbl_top = tbls[tbl_index][1][0][3]
tbl_bottom = tbls[tbl_index][1][0][4]
tbl_tag = "@@{}\t{:.1f}\t{:.1f}\t{:.1f}\t{:.1f}##" \
.format(tbl_pn, tbl_left, tbl_right, tbl_top, tbl_bottom)
tbl_text = ''.join(tbls[tbl_index][0][1])
return tbl_pn, tbl_left, tbl_right, tbl_top, tbl_bottom, tbl_tag, tbl_text
class Docx(DocxParser):
def __init__(self):
pass
def get_picture(self, document, paragraph):
img = paragraph._element.xpath('.//pic:pic')
if not img:
return None
img = img[0]
embed = img.xpath('.//a:blip/@r:embed')[0]
related_part = document.part.related_parts[embed]
image = related_part.image
image = Image.open(BytesIO(image.blob)).convert('RGB')
return image
def __call__(self, filename, binary=None, from_page=0, to_page=100000, callback=None):
self.doc = Document(
filename) if not binary else Document(BytesIO(binary))
pn = 0
last_answer, last_image = "", None
question_stack, level_stack = [], []
qai_list = []
for p in self.doc.paragraphs:
if pn > to_page:
break
question_level, p_text = 0, ''
if from_page <= pn < to_page and p.text.strip():
question_level, p_text = docx_question_level(p)
if not question_level or question_level > 6: # not a question
last_answer = f'{last_answer}\n{p_text}'
current_image = self.get_picture(self.doc, p)
last_image = concat_img(last_image, current_image)
else: # is a question
if last_answer or last_image:
sum_question = '\n'.join(question_stack)
if sum_question:
qai_list.append((sum_question, last_answer, last_image))
last_answer, last_image = '', None
i = question_level
while question_stack and i <= level_stack[-1]:
question_stack.pop()
level_stack.pop()
question_stack.append(p_text)
level_stack.append(question_level)
for run in p.runs:
if 'lastRenderedPageBreak' in run._element.xml:
pn += 1
continue
if 'w:br' in run._element.xml and 'type="page"' in run._element.xml:
pn += 1
if last_answer:
sum_question = '\n'.join(question_stack)
if sum_question:
qai_list.append((sum_question, last_answer, last_image))
tbls = []
for tb in self.doc.tables:
html= "<table>"
for r in tb.rows:
html += "<tr>"
i = 0
while i < len(r.cells):
span = 1
c = r.cells[i]
for j in range(i+1, len(r.cells)):
if c.text == r.cells[j].text:
span += 1
i = j
i += 1
html += f"<td>{c.text}</td>" if span == 1 else f"<td colspan='{span}'>{c.text}</td>"
html += "</tr>"
html += "</table>"
tbls.append(((None, html), ""))
return qai_list, tbls
def rmPrefix(txt):
return re.sub(
r"^(问题|答案|回答|user|assistant|Q|A|Question|Answer|问|答)[\t:： ]+", "", txt.strip(), flags=re.IGNORECASE)
def beAdocPdf(d, q, a, eng, image, poss):
qprefix = "Question: " if eng else "问题："
aprefix = "Answer: " if eng else "回答："
d["content_with_weight"] = "\t".join(
[qprefix + rmPrefix(q), aprefix + rmPrefix(a)])
d["content_ltks"] = rag_tokenizer.tokenize(q)
d["content_sm_ltks"] = rag_tokenizer.fine_grained_tokenize(d["content_ltks"])
d["image"] = image
add_positions(d, poss)
return d
def beAdocDocx(d, q, a, eng, image):
qprefix = "Question: " if eng else "问题："
aprefix = "Answer: " if eng else "回答："
d["content_with_weight"] = "\t".join(
[qprefix + rmPrefix(q), aprefix + rmPrefix(a)])
d["content_ltks"] = rag_tokenizer.tokenize(q)
d["content_sm_ltks"] = rag_tokenizer.fine_grained_tokenize(d["content_ltks"])
d["image"] = image
return d
def beAdoc(d, q, a, eng):
qprefix = "Question: " if eng else "问题："
aprefix = "Answer: " if eng else "回答："
d["content_with_weight"] = "\t".join(
[qprefix + rmPrefix(q), aprefix + rmPrefix(a)])
d["content_ltks"] = rag_tokenizer.tokenize(q)
d["content_sm_ltks"] = rag_tokenizer.fine_grained_tokenize(d["content_ltks"])
return d
def mdQuestionLevel(s):
match = re.match(r'#*', s)
return (len(match.group(0)), s.lstrip('#').lstrip()) if match else (0, s)
def chunk(filename, binary=None, lang="Chinese", callback=None, **kwargs):
"""
Excel and csv(txt) format files are supported.
If the file is in excel format, there should be 2 column question and answer without header.
And question column is ahead of answer column.
And it's O.K if it has multiple sheets as long as the columns are rightly composed.
If it's in csv format, it should be UTF-8 encoded. Use TAB as delimiter to separate question and answer.
All the deformed lines will be ignored.
Every pair of Q&A will be treated as a chunk.
"""
eng = lang.lower() == "english"
res = []
doc = {
"docnm_kwd": filename,
"title_tks": rag_tokenizer.tokenize(re.sub(r"\.[a-zA-Z]+$", "", filename))
}
if re.search(r"\.xlsx?$", filename, re.IGNORECASE):
callback(0.1, "Start to parse.")
excel_parser = Excel()
for q, a in excel_parser(filename, binary, callback):
res.append(beAdoc(deepcopy(doc), q, a, eng))
return res
elif re.search(r"\.(txt|csv)$", filename, re.IGNORECASE):
callback(0.1, "Start to parse.")
txt = ""
if binary:
encoding = find_codec(binary)
txt = binary.decode(encoding, errors="ignore")
else:
with open(filename, "r") as f:
while True:
l = f.readline()
if not l:
break
txt += l
lines = txt.split("\n")
comma, tab = 0, 0
for l in lines:
if len(l.split(",")) == 2: comma += 1
if len(l.split("\t")) == 2: tab += 1
delimiter = "\t" if tab >= comma else ","
fails = []
question, answer = "", ""
i = 0
while i < len(lines):
arr = lines[i].split(delimiter)
if len(arr) != 2:
if question: answer += "\n" + lines[i]
else:
fails.append(str(i+1))
elif len(arr) == 2:
if question and answer: res.append(beAdoc(deepcopy(doc), question, answer, eng))
question, answer = arr
i += 1
if len(res) % 999 == 0:
callback(len(res) * 0.6 / len(lines), ("Extract Q&A: {}".format(len(res)) + (
f"{len(fails)} failure, line: %s..." % (",".join(fails[:3])) if fails else "")))
if question: res.append(beAdoc(deepcopy(doc), question, answer, eng))
callback(0.6, ("Extract Q&A: {}".format(len(res)) + (
f"{len(fails)} failure, line: %s..." % (",".join(fails[:3])) if fails else "")))
return res
elif re.search(r"\.pdf$", filename, re.IGNORECASE):
callback(0.1, "Start to parse.")
pdf_parser = Pdf()
qai_list, tbls = pdf_parser(filename if not binary else binary,
from_page=0, to_page=10000, callback=callback)
for q, a, image, poss in qai_list:
res.append(beAdocPdf(deepcopy(doc), q, a, eng, image, poss))
return res
elif re.search(r"\.(md|markdown)$", filename, re.IGNORECASE):
callback(0.1, "Start to parse.")
txt = ""
if binary:
encoding = find_codec(binary)
txt = binary.decode(encoding, errors="ignore")
else:
with open(filename, "r") as f:
while True:
l = f.readline()
if not l:
break
txt += l
lines = txt.split("\n")
last_question, last_answer = "", ""
question_stack, level_stack = [], []
code_block = False
level_index = [-1] * 7
for index, l in enumerate(lines):
if l.strip().startswith('```'):
code_block = not code_block
question_level, question = 0, ''
if not code_block:
question_level, question = mdQuestionLevel(l)
if not question_level or question_level > 6: # not a question
last_answer = f'{last_answer}\n{l}'
else: # is a question
if last_answer.strip():
sum_question = '\n'.join(question_stack)
if sum_question:
res.append(beAdoc(deepcopy(doc), sum_question, markdown(last_answer, extensions=['markdown.extensions.tables']), eng))
last_answer = ''
i = question_level
while question_stack and i <= level_stack[-1]:
question_stack.pop()
level_stack.pop()
question_stack.append(question)
level_stack.append(question_level)
if last_answer.strip():
sum_question = '\n'.join(question_stack)
if sum_question:
res.append(beAdoc(deepcopy(doc), sum_question, markdown(last_answer, extensions=['markdown.extensions.tables']), eng))
return res
elif re.search(r"\.docx$", filename, re.IGNORECASE):
docx_parser = Docx()
qai_list, tbls = docx_parser(filename, binary,
from_page=0, to_page=10000, callback=callback)
res = tokenize_table(tbls, doc, eng)
for q, a, image in qai_list:
res.append(beAdocDocx(deepcopy(doc), q, a, eng, image))
return res
raise NotImplementedError(
"Excel, csv(txt), pdf, markdown and docx format files are supported.")
if __name__ == "__main__":
import sys
def dummy(prog=None, msg=""):
pass
chunk(sys.argv[1], from_page=0, to_page=10, callback=dummy)

346
rag/app/resume.py
View File

@ -1,173 +1,173 @@
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
#
import base64
import datetime
import json
import re
import pandas as pd
import requests
from api.db.services.knowledgebase_service import KnowledgebaseService
from rag.nlp import rag_tokenizer
from deepdoc.parser.resume import refactor
from deepdoc.parser.resume import step_one, step_two
from rag.settings import cron_logger
from rag.utils import rmSpace
forbidden_select_fields4resume = [
"name_pinyin_kwd", "edu_first_fea_kwd", "degree_kwd", "sch_rank_kwd", "edu_fea_kwd"
]
def remote_call(filename, binary):
q = {
"header": {
"uid": 1,
"user": "kevinhu",
"log_id": filename
},
"request": {
"p": {
"request_id": "1",
"encrypt_type": "base64",
"filename": filename,
"langtype": '',
"fileori": base64.b64encode(binary).decode('utf-8')
},
"c": "resume_parse_module",
"m": "resume_parse"
}
}
for _ in range(3):
try:
resume = requests.post(
"http://127.0.0.1:61670/tog",
data=json.dumps(q))
resume = resume.json()["response"]["results"]
resume = refactor(resume)
for k in ["education", "work", "project",
"training", "skill", "certificate", "language"]:
if not resume.get(k) and k in resume:
del resume[k]
resume = step_one.refactor(pd.DataFrame([{"resume_content": json.dumps(resume), "tob_resume_id": "x",
"updated_at": datetime.datetime.now().strftime("%Y-%m-%d %H:%M:%S")}]))
resume = step_two.parse(resume)
return resume
except Exception as e:
cron_logger.error("Resume parser error: " + str(e))
return {}
def chunk(filename, binary=None, callback=None, **kwargs):
"""
The supported file formats are pdf, docx and txt.
To maximize the effectiveness, parse the resume correctly, please concat us: https://github.com/infiniflow/ragflow
"""
if not re.search(r"\.(pdf|doc|docx|txt)$", filename, flags=re.IGNORECASE):
raise NotImplementedError("file type not supported yet(pdf supported)")
if not binary:
with open(filename, "rb") as f:
binary = f.read()
callback(0.2, "Resume parsing is going on...")
resume = remote_call(filename, binary)
if len(resume.keys()) < 7:
callback(-1, "Resume is not successfully parsed.")
raise Exception("Resume parser remote call fail!")
callback(0.6, "Done parsing. Chunking...")
print(json.dumps(resume, ensure_ascii=False, indent=2))
field_map = {
"name_kwd": "姓名/名字",
"name_pinyin_kwd": "姓名拼音/名字拼音",
"gender_kwd": "性别（男，女）",
"age_int": "年龄/岁/年纪",
"phone_kwd": "电话/手机/微信",
"email_tks": "email/e-mail/邮箱",
"position_name_tks": "职位/职能/岗位/职责",
"expect_city_names_tks": "期望城市",
"work_exp_flt": "工作年限/工作年份/N年经验/毕业了多少年",
"corporation_name_tks": "最近就职(上班)的公司/上一家公司",
"first_school_name_tks": "第一学历毕业学校",
"first_degree_kwd": "第一学历（高中，职高，硕士，本科，博士，初中，中技，中专，专科，专升本，MPA，MBA，EMBA）",
"highest_degree_kwd": "最高学历（高中，职高，硕士，本科，博士，初中，中技，中专，专科，专升本，MPA，MBA，EMBA）",
"first_major_tks": "第一学历专业",
"edu_first_fea_kwd": "第一学历标签（211，留学，双一流，985，海外知名，重点大学，中专，专升本，专科，本科，大专）",
"degree_kwd": "过往学历（高中，职高，硕士，本科，博士，初中，中技，中专，专科，专升本，MPA，MBA，EMBA）",
"major_tks": "学过的专业/过往专业",
"school_name_tks": "学校/毕业院校",
"sch_rank_kwd": "学校标签（顶尖学校，精英学校，优质学校，一般学校）",
"edu_fea_kwd": "教育标签（211，留学，双一流，985，海外知名，重点大学，中专，专升本，专科，本科，大专）",
"corp_nm_tks": "就职过的公司/之前的公司/上过班的公司",
"edu_end_int": "毕业年份",
"industry_name_tks": "所在行业",
"birth_dt": "生日/出生年份",
"expect_position_name_tks": "期望职位/期望职能/期望岗位",
}
titles = []
for n in ["name_kwd", "gender_kwd", "position_name_tks", "age_int"]:
v = resume.get(n, "")
if isinstance(v, list):
v = v[0]
if n.find("tks") > 0:
v = rmSpace(v)
titles.append(str(v))
doc = {
"docnm_kwd": filename,
"title_tks": rag_tokenizer.tokenize("-".join(titles) + "-简历")
}
doc["title_sm_tks"] = rag_tokenizer.fine_grained_tokenize(doc["title_tks"])
pairs = []
for n, m in field_map.items():
if not resume.get(n):
continue
v = resume[n]
if isinstance(v, list):
v = " ".join(v)
if n.find("tks") > 0:
v = rmSpace(v)
pairs.append((m, str(v)))
doc["content_with_weight"] = "\n".join(
["{}: {}".format(re.sub(r"（[^（）]+）", "", k), v) for k, v in pairs])
doc["content_ltks"] = rag_tokenizer.tokenize(doc["content_with_weight"])
doc["content_sm_ltks"] = rag_tokenizer.fine_grained_tokenize(doc["content_ltks"])
for n, _ in field_map.items():
if n not in resume:
continue
if isinstance(resume[n], list) and (
len(resume[n]) == 1 or n not in forbidden_select_fields4resume):
resume[n] = resume[n][0]
if n.find("_tks") > 0:
resume[n] = rag_tokenizer.fine_grained_tokenize(resume[n])
doc[n] = resume[n]
print(doc)
KnowledgebaseService.update_parser_config(
kwargs["kb_id"], {"field_map": field_map})
return [doc]
if __name__ == "__main__":
import sys
def dummy(a, b):
pass
chunk(sys.argv[1], callback=dummy)
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
#
import base64
import datetime
import json
import re
import pandas as pd
import requests
from api.db.services.knowledgebase_service import KnowledgebaseService
from rag.nlp import rag_tokenizer
from deepdoc.parser.resume import refactor
from deepdoc.parser.resume import step_one, step_two
from rag.settings import cron_logger
from rag.utils import rmSpace
forbidden_select_fields4resume = [
"name_pinyin_kwd", "edu_first_fea_kwd", "degree_kwd", "sch_rank_kwd", "edu_fea_kwd"
]
def remote_call(filename, binary):
q = {
"header": {
"uid": 1,
"user": "kevinhu",
"log_id": filename
},
"request": {
"p": {
"request_id": "1",
"encrypt_type": "base64",
"filename": filename,
"langtype": '',
"fileori": base64.b64encode(binary).decode('utf-8')
},
"c": "resume_parse_module",
"m": "resume_parse"
}
}
for _ in range(3):
try:
resume = requests.post(
"http://127.0.0.1:61670/tog",
data=json.dumps(q))
resume = resume.json()["response"]["results"]
resume = refactor(resume)
for k in ["education", "work", "project",
"training", "skill", "certificate", "language"]:
if not resume.get(k) and k in resume:
del resume[k]
resume = step_one.refactor(pd.DataFrame([{"resume_content": json.dumps(resume), "tob_resume_id": "x",
"updated_at": datetime.datetime.now().strftime("%Y-%m-%d %H:%M:%S")}]))
resume = step_two.parse(resume)
return resume
except Exception as e:
cron_logger.error("Resume parser error: " + str(e))
return {}
def chunk(filename, binary=None, callback=None, **kwargs):
"""
The supported file formats are pdf, docx and txt.
To maximize the effectiveness, parse the resume correctly, please concat us: https://github.com/infiniflow/ragflow
"""
if not re.search(r"\.(pdf|doc|docx|txt)$", filename, flags=re.IGNORECASE):
raise NotImplementedError("file type not supported yet(pdf supported)")
if not binary:
with open(filename, "rb") as f:
binary = f.read()
callback(0.2, "Resume parsing is going on...")
resume = remote_call(filename, binary)
if len(resume.keys()) < 7:
callback(-1, "Resume is not successfully parsed.")
raise Exception("Resume parser remote call fail!")
callback(0.6, "Done parsing. Chunking...")
print(json.dumps(resume, ensure_ascii=False, indent=2))
field_map = {
"name_kwd": "姓名/名字",
"name_pinyin_kwd": "姓名拼音/名字拼音",
"gender_kwd": "性别（男，女）",
"age_int": "年龄/岁/年纪",
"phone_kwd": "电话/手机/微信",
"email_tks": "email/e-mail/邮箱",
"position_name_tks": "职位/职能/岗位/职责",
"expect_city_names_tks": "期望城市",
"work_exp_flt": "工作年限/工作年份/N年经验/毕业了多少年",
"corporation_name_tks": "最近就职(上班)的公司/上一家公司",
"first_school_name_tks": "第一学历毕业学校",
"first_degree_kwd": "第一学历（高中，职高，硕士，本科，博士，初中，中技，中专，专科，专升本，MPA，MBA，EMBA）",
"highest_degree_kwd": "最高学历（高中，职高，硕士，本科，博士，初中，中技，中专，专科，专升本，MPA，MBA，EMBA）",
"first_major_tks": "第一学历专业",
"edu_first_fea_kwd": "第一学历标签（211，留学，双一流，985，海外知名，重点大学，中专，专升本，专科，本科，大专）",
"degree_kwd": "过往学历（高中，职高，硕士，本科，博士，初中，中技，中专，专科，专升本，MPA，MBA，EMBA）",
"major_tks": "学过的专业/过往专业",
"school_name_tks": "学校/毕业院校",
"sch_rank_kwd": "学校标签（顶尖学校，精英学校，优质学校，一般学校）",
"edu_fea_kwd": "教育标签（211，留学，双一流，985，海外知名，重点大学，中专，专升本，专科，本科，大专）",
"corp_nm_tks": "就职过的公司/之前的公司/上过班的公司",
"edu_end_int": "毕业年份",
"industry_name_tks": "所在行业",
"birth_dt": "生日/出生年份",
"expect_position_name_tks": "期望职位/期望职能/期望岗位",
}
titles = []
for n in ["name_kwd", "gender_kwd", "position_name_tks", "age_int"]:
v = resume.get(n, "")
if isinstance(v, list):
v = v[0]
if n.find("tks") > 0:
v = rmSpace(v)
titles.append(str(v))
doc = {
"docnm_kwd": filename,
"title_tks": rag_tokenizer.tokenize("-".join(titles) + "-简历")
}
doc["title_sm_tks"] = rag_tokenizer.fine_grained_tokenize(doc["title_tks"])
pairs = []
for n, m in field_map.items():
if not resume.get(n):
continue
v = resume[n]
if isinstance(v, list):
v = " ".join(v)
if n.find("tks") > 0:
v = rmSpace(v)
pairs.append((m, str(v)))
doc["content_with_weight"] = "\n".join(
["{}: {}".format(re.sub(r"（[^（）]+）", "", k), v) for k, v in pairs])
doc["content_ltks"] = rag_tokenizer.tokenize(doc["content_with_weight"])
doc["content_sm_ltks"] = rag_tokenizer.fine_grained_tokenize(doc["content_ltks"])
for n, _ in field_map.items():
if n not in resume:
continue
if isinstance(resume[n], list) and (
len(resume[n]) == 1 or n not in forbidden_select_fields4resume):
resume[n] = resume[n][0]
if n.find("_tks") > 0:
resume[n] = rag_tokenizer.fine_grained_tokenize(resume[n])
doc[n] = resume[n]
print(doc)
KnowledgebaseService.update_parser_config(
kwargs["kb_id"], {"field_map": field_map})
return [doc]
if __name__ == "__main__":
import sys
def dummy(a, b):
pass
chunk(sys.argv[1], callback=dummy)

504
rag/app/table.py
View File

@ -1,252 +1,252 @@
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
#
import copy
import re
from io import BytesIO
from xpinyin import Pinyin
import numpy as np
import pandas as pd
from openpyxl import load_workbook
from dateutil.parser import parse as datetime_parse
from api.db.services.knowledgebase_service import KnowledgebaseService
from rag.nlp import rag_tokenizer, is_english, tokenize, find_codec
from deepdoc.parser import ExcelParser
class Excel(ExcelParser):
def __call__(self, fnm, binary=None, from_page=0,
to_page=10000000000, callback=None):
if not binary:
wb = load_workbook(fnm)
else:
wb = load_workbook(BytesIO(binary))
total = 0
for sheetname in wb.sheetnames:
total += len(list(wb[sheetname].rows))
res, fails, done = [], [], 0
rn = 0
for sheetname in wb.sheetnames:
ws = wb[sheetname]
rows = list(ws.rows)
if not rows:continue
headers = [cell.value for cell in rows[0]]
missed = set([i for i, h in enumerate(headers) if h is None])
headers = [
cell.value for i,
cell in enumerate(
rows[0]) if i not in missed]
if not headers:continue
data = []
for i, r in enumerate(rows[1:]):
rn += 1
if rn - 1 < from_page:
continue
if rn - 1 >= to_page:
break
row = [
cell.value for ii,
cell in enumerate(r) if ii not in missed]
if len(row) != len(headers):
fails.append(str(i))
continue
data.append(row)
done += 1
res.append(pd.DataFrame(np.array(data), columns=headers))
callback(0.3, ("Extract records: {}~{}".format(from_page + 1, min(to_page, from_page + rn)) + (
f"{len(fails)} failure, line: %s..." % (",".join(fails[:3])) if fails else "")))
return res
def trans_datatime(s):
try:
return datetime_parse(s.strip()).strftime("%Y-%m-%d %H:%M:%S")
except Exception as e:
pass
def trans_bool(s):
if re.match(r"(true|yes|是|\*|✓|✔|☑|✅|√)$",
str(s).strip(), flags=re.IGNORECASE):
return "yes"
if re.match(r"(false|no|否|⍻|×)$", str(s).strip(), flags=re.IGNORECASE):
return "no"
def column_data_type(arr):
arr = list(arr)
uni = len(set([a for a in arr if a is not None]))
counts = {"int": 0, "float": 0, "text": 0, "datetime": 0, "bool": 0}
trans = {t: f for f, t in
[(int, "int"), (float, "float"), (trans_datatime, "datetime"), (trans_bool, "bool"), (str, "text")]}
for a in arr:
if a is None:
continue
if re.match(r"[+-]?[0-9]+(\.0+)?$", str(a).replace("%%", "")):
counts["int"] += 1
elif re.match(r"[+-]?[0-9.]+$", str(a).replace("%%", "")):
counts["float"] += 1
elif re.match(r"(true|yes|是|\*|✓|✔|☑|✅|√|false|no|否|⍻|×)$", str(a), flags=re.IGNORECASE):
counts["bool"] += 1
elif trans_datatime(str(a)):
counts["datetime"] += 1
else:
counts["text"] += 1
counts = sorted(counts.items(), key=lambda x: x[1] * -1)
ty = counts[0][0]
for i in range(len(arr)):
if arr[i] is None:
continue
try:
arr[i] = trans[ty](str(arr[i]))
except Exception as e:
arr[i] = None
# if ty == "text":
# if len(arr) > 128 and uni / len(arr) < 0.1:
# ty = "keyword"
return arr, ty
def chunk(filename, binary=None, from_page=0, to_page=10000000000,
lang="Chinese", callback=None, **kwargs):
"""
Excel and csv(txt) format files are supported.
For csv or txt file, the delimiter between columns is TAB.
The first line must be column headers.
Column headers must be meaningful terms inorder to make our NLP model understanding.
It's good to enumerate some synonyms using slash '/' to separate, and even better to
enumerate values using brackets like 'gender/sex(male, female)'.
Here are some examples for headers:
1. supplier/vendor\tcolor(yellow, red, brown)\tgender/sex(male, female)\tsize(M,L,XL,XXL)
2. 姓名/名字\t电话/手机/微信\t最高学历（高中，职高，硕士，本科，博士，初中，中技，中专，专科，专升本，MPA，MBA，EMBA）
Every row in table will be treated as a chunk.
"""
if re.search(r"\.xlsx?$", filename, re.IGNORECASE):
callback(0.1, "Start to parse.")
excel_parser = Excel()
dfs = excel_parser(
filename,
binary,
from_page=from_page,
to_page=to_page,
callback=callback)
elif re.search(r"\.(txt|csv)$", filename, re.IGNORECASE):
callback(0.1, "Start to parse.")
txt = ""
if binary:
encoding = find_codec(binary)
txt = binary.decode(encoding, errors="ignore")
else:
with open(filename, "r") as f:
while True:
l = f.readline()
if not l:
break
txt += l
lines = txt.split("\n")
fails = []
headers = lines[0].split(kwargs.get("delimiter", "\t"))
rows = []
for i, line in enumerate(lines[1:]):
if i < from_page:
continue
if i >= to_page:
break
row = [l for l in line.split(kwargs.get("delimiter", "\t"))]
if len(row) != len(headers):
fails.append(str(i))
continue
rows.append(row)
callback(0.3, ("Extract records: {}~{}".format(from_page, min(len(lines), to_page)) + (
f"{len(fails)} failure, line: %s..." % (",".join(fails[:3])) if fails else "")))
dfs = [pd.DataFrame(np.array(rows), columns=headers)]
else:
raise NotImplementedError(
"file type not supported yet(excel, text, csv supported)")
res = []
PY = Pinyin()
fieds_map = {
"text": "_tks",
"int": "_long",
"keyword": "_kwd",
"float": "_flt",
"datetime": "_dt",
"bool": "_kwd"}
for df in dfs:
for n in ["id", "_id", "index", "idx"]:
if n in df.columns:
del df[n]
clmns = df.columns.values
txts = list(copy.deepcopy(clmns))
py_clmns = [
PY.get_pinyins(
re.sub(
r"(/.*|（[^（）]+?）|\([^()]+?\))",
"",
str(n)),
'_')[0] for n in clmns]
clmn_tys = []
for j in range(len(clmns)):
cln, ty = column_data_type(df[clmns[j]])
clmn_tys.append(ty)
df[clmns[j]] = cln
if ty == "text":
txts.extend([str(c) for c in cln if c])
clmns_map = [(py_clmns[i].lower() + fieds_map[clmn_tys[i]], str(clmns[i]).replace("_", " "))
for i in range(len(clmns))]
eng = lang.lower() == "english" # is_english(txts)
for ii, row in df.iterrows():
d = {
"docnm_kwd": filename,
"title_tks": rag_tokenizer.tokenize(re.sub(r"\.[a-zA-Z]+$", "", filename))
}
row_txt = []
for j in range(len(clmns)):
if row[clmns[j]] is None:
continue
if not str(row[clmns[j]]):
continue
if pd.isna(row[clmns[j]]):
continue
fld = clmns_map[j][0]
d[fld] = row[clmns[j]] if clmn_tys[j] != "text" else rag_tokenizer.tokenize(
row[clmns[j]])
row_txt.append("{}:{}".format(clmns[j], row[clmns[j]]))
if not row_txt:
continue
tokenize(d, "; ".join(row_txt), eng)
res.append(d)
KnowledgebaseService.update_parser_config(
kwargs["kb_id"], {"field_map": {k: v for k, v in clmns_map}})
callback(0.35, "")
return res
if __name__ == "__main__":
import sys
def dummy(prog=None, msg=""):
pass
chunk(sys.argv[1], callback=dummy)
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
#
import copy
import re
from io import BytesIO
from xpinyin import Pinyin
import numpy as np
import pandas as pd
from openpyxl import load_workbook
from dateutil.parser import parse as datetime_parse
from api.db.services.knowledgebase_service import KnowledgebaseService
from rag.nlp import rag_tokenizer, is_english, tokenize, find_codec
from deepdoc.parser import ExcelParser
class Excel(ExcelParser):
def __call__(self, fnm, binary=None, from_page=0,
to_page=10000000000, callback=None):
if not binary:
wb = load_workbook(fnm)
else:
wb = load_workbook(BytesIO(binary))
total = 0
for sheetname in wb.sheetnames:
total += len(list(wb[sheetname].rows))
res, fails, done = [], [], 0
rn = 0
for sheetname in wb.sheetnames:
ws = wb[sheetname]
rows = list(ws.rows)
if not rows:continue
headers = [cell.value for cell in rows[0]]
missed = set([i for i, h in enumerate(headers) if h is None])
headers = [
cell.value for i,
cell in enumerate(
rows[0]) if i not in missed]
if not headers:continue
data = []
for i, r in enumerate(rows[1:]):
rn += 1
if rn - 1 < from_page:
continue
if rn - 1 >= to_page:
break
row = [
cell.value for ii,
cell in enumerate(r) if ii not in missed]
if len(row) != len(headers):
fails.append(str(i))
continue
data.append(row)
done += 1
res.append(pd.DataFrame(np.array(data), columns=headers))
callback(0.3, ("Extract records: {}~{}".format(from_page + 1, min(to_page, from_page + rn)) + (
f"{len(fails)} failure, line: %s..." % (",".join(fails[:3])) if fails else "")))
return res
def trans_datatime(s):
try:
return datetime_parse(s.strip()).strftime("%Y-%m-%d %H:%M:%S")
except Exception as e:
pass
def trans_bool(s):
if re.match(r"(true|yes|是|\*|✓|✔|☑|✅|√)$",
str(s).strip(), flags=re.IGNORECASE):
return "yes"
if re.match(r"(false|no|否|⍻|×)$", str(s).strip(), flags=re.IGNORECASE):
return "no"
def column_data_type(arr):
arr = list(arr)
uni = len(set([a for a in arr if a is not None]))
counts = {"int": 0, "float": 0, "text": 0, "datetime": 0, "bool": 0}
trans = {t: f for f, t in
[(int, "int"), (float, "float"), (trans_datatime, "datetime"), (trans_bool, "bool"), (str, "text")]}
for a in arr:
if a is None:
continue
if re.match(r"[+-]?[0-9]+(\.0+)?$", str(a).replace("%%", "")):
counts["int"] += 1
elif re.match(r"[+-]?[0-9.]+$", str(a).replace("%%", "")):
counts["float"] += 1
elif re.match(r"(true|yes|是|\*|✓|✔|☑|✅|√|false|no|否|⍻|×)$", str(a), flags=re.IGNORECASE):
counts["bool"] += 1
elif trans_datatime(str(a)):
counts["datetime"] += 1
else:
counts["text"] += 1
counts = sorted(counts.items(), key=lambda x: x[1] * -1)
ty = counts[0][0]
for i in range(len(arr)):
if arr[i] is None:
continue
try:
arr[i] = trans[ty](str(arr[i]))
except Exception as e:
arr[i] = None
# if ty == "text":
# if len(arr) > 128 and uni / len(arr) < 0.1:
# ty = "keyword"
return arr, ty
def chunk(filename, binary=None, from_page=0, to_page=10000000000,
lang="Chinese", callback=None, **kwargs):
"""
Excel and csv(txt) format files are supported.
For csv or txt file, the delimiter between columns is TAB.
The first line must be column headers.
Column headers must be meaningful terms inorder to make our NLP model understanding.
It's good to enumerate some synonyms using slash '/' to separate, and even better to
enumerate values using brackets like 'gender/sex(male, female)'.
Here are some examples for headers:
1. supplier/vendor\tcolor(yellow, red, brown)\tgender/sex(male, female)\tsize(M,L,XL,XXL)
2. 姓名/名字\t电话/手机/微信\t最高学历（高中，职高，硕士，本科，博士，初中，中技，中专，专科，专升本，MPA，MBA，EMBA）
Every row in table will be treated as a chunk.
"""
if re.search(r"\.xlsx?$", filename, re.IGNORECASE):
callback(0.1, "Start to parse.")
excel_parser = Excel()
dfs = excel_parser(
filename,
binary,
from_page=from_page,
to_page=to_page,
callback=callback)
elif re.search(r"\.(txt|csv)$", filename, re.IGNORECASE):
callback(0.1, "Start to parse.")
txt = ""
if binary:
encoding = find_codec(binary)
txt = binary.decode(encoding, errors="ignore")
else:
with open(filename, "r") as f:
while True:
l = f.readline()
if not l:
break
txt += l
lines = txt.split("\n")
fails = []
headers = lines[0].split(kwargs.get("delimiter", "\t"))
rows = []
for i, line in enumerate(lines[1:]):
if i < from_page:
continue
if i >= to_page:
break
row = [l for l in line.split(kwargs.get("delimiter", "\t"))]
if len(row) != len(headers):
fails.append(str(i))
continue
rows.append(row)
callback(0.3, ("Extract records: {}~{}".format(from_page, min(len(lines), to_page)) + (
f"{len(fails)} failure, line: %s..." % (",".join(fails[:3])) if fails else "")))
dfs = [pd.DataFrame(np.array(rows), columns=headers)]
else:
raise NotImplementedError(
"file type not supported yet(excel, text, csv supported)")
res = []
PY = Pinyin()
fieds_map = {
"text": "_tks",
"int": "_long",
"keyword": "_kwd",
"float": "_flt",
"datetime": "_dt",
"bool": "_kwd"}
for df in dfs:
for n in ["id", "_id", "index", "idx"]:
if n in df.columns:
del df[n]
clmns = df.columns.values
txts = list(copy.deepcopy(clmns))
py_clmns = [
PY.get_pinyins(
re.sub(
r"(/.*|（[^（）]+?）|\([^()]+?\))",
"",
str(n)),
'_')[0] for n in clmns]
clmn_tys = []
for j in range(len(clmns)):
cln, ty = column_data_type(df[clmns[j]])
clmn_tys.append(ty)
df[clmns[j]] = cln
if ty == "text":
txts.extend([str(c) for c in cln if c])
clmns_map = [(py_clmns[i].lower() + fieds_map[clmn_tys[i]], str(clmns[i]).replace("_", " "))
for i in range(len(clmns))]
eng = lang.lower() == "english" # is_english(txts)
for ii, row in df.iterrows():
d = {
"docnm_kwd": filename,
"title_tks": rag_tokenizer.tokenize(re.sub(r"\.[a-zA-Z]+$", "", filename))
}
row_txt = []
for j in range(len(clmns)):
if row[clmns[j]] is None:
continue
if not str(row[clmns[j]]):
continue
if pd.isna(row[clmns[j]]):
continue
fld = clmns_map[j][0]
d[fld] = row[clmns[j]] if clmn_tys[j] != "text" else rag_tokenizer.tokenize(
row[clmns[j]])
row_txt.append("{}:{}".format(clmns[j], row[clmns[j]]))
if not row_txt:
continue
tokenize(d, "; ".join(row_txt), eng)
res.append(d)
KnowledgebaseService.update_parser_config(
kwargs["kb_id"], {"field_map": {k: v for k, v in clmns_map}})
callback(0.35, "")
return res
if __name__ == "__main__":
import sys
def dummy(prog=None, msg=""):
pass
chunk(sys.argv[1], callback=dummy)